fix: merge worker restore into develop host lifecycle

Integrate integration/worker-restore without reviving removed Worker socket/discovery paths. Keep restore candidate allocation and canonical effects deferred until activation, with registry reclamation and observer publication fences on the in-process controller.

Preserve develop BackendJob/provenance, immutable Journal body capture and erasure, and Workdir diagnostics. Adapt obsolete socket fault tests to in-process publication and registry drift boundaries.

Validation: root cargo check; ordinary worker, worker-runtime and standalone package tests; Server restore (43) and Journal (49) filters; mcp/server-api/runtime-api/session-store tests; optional Runtime feature check; OpenAPI and six generated TypeScript checks; Web restore unit tests (15); runner tests (21); fmt, naming and diff checks. Full workspace/Server suites, E2E, live/model and fixture generation not run.

crates/mcp/src/stdio.rs

@@ -617,7 +617,8 @@ impl fmt::Display for McpRequestStage {
     }
 }
 
-/// A running initialized stdio MCP client.
+/// An owned stdio MCP process, initialized by `connect` or explicitly after
+/// `spawn_uninitialized` when startup cleanup must retain it before any await.
 pub struct McpStdioClient {
     server_name: String,
     limits: McpStdioLimits,
@@ -636,6 +637,9 @@ pub struct McpStdioClient {
 
 impl McpStdioClient {
     /// Spawn, initialize, negotiate capabilities, and send notifications/initialized.
+    /// Initialization errors attempt shutdown, but this convenience API cannot
+    /// return an unreaped client on failure/cancellation. Its error is not proof
+    /// of cleanup; use `spawn_uninitialized` with retained ownership when needed.
     pub async fn connect(
         spec: McpStdioServerSpec,
         limits: McpStdioLimits,
@@ -667,6 +671,21 @@ impl McpStdioClient {
     async fn spawn(
         spec: McpStdioServerSpec,
         limits: McpStdioLimits,
+    ) -> Result<Self, McpClientError> {
+        Self::spawn_uninitialized(spec, limits)
+    }
+
+    /// Spawn without awaiting the handshake, so a cleanup owner can retain the
+    /// client before initialization can fail or be cancelled. Call
+    /// `initialize_with_timeout` before discovery; on error/cancellation, shut
+    /// down this same client rather than spawning a replacement.
+    ///
+    /// Once the process spawns, this returns its owner even if an expected pipe
+    /// is missing. Initialization then fails, but cleanup can still reap the
+    /// actual child rather than relying on kill-on-drop for a post-spawn error.
+    pub fn spawn_uninitialized(
+        spec: McpStdioServerSpec,
+        limits: McpStdioLimits,
     ) -> Result<Self, McpClientError> {
         let redactor = spec.redactor();
         let mut command = Command::new(&spec.command);
@@ -688,29 +707,23 @@ impl McpStdioClient {
                 McpErrorKind::Io(redactor.redact(&err.to_string())),
             )
         })?;
-        let stdin = child.stdin.take().ok_or_else(|| {
-            McpClientError::new(
-                &spec.name,
-                McpPhase::Spawn,
-                McpErrorKind::Protocol("child stdin was not piped".into()),
-            )
-        })?;
-        let stdout = child.stdout.take().ok_or_else(|| {
-            McpClientError::new(
-                &spec.name,
-                McpPhase::Spawn,
-                McpErrorKind::Protocol("child stdout was not piped".into()),
-            )
-        })?;
-        let stderr = child.stderr.take().ok_or_else(|| {
-            McpClientError::new(
-                &spec.name,
-                McpPhase::Spawn,
-                McpErrorKind::Protocol("child stderr was not piped".into()),
-            )
-        })?;
+        let stdin = child.stdin.take();
+        let stdout = child.stdout.take();
+        let stderr = child.stderr.take();
+        let missing_pipe = if stdin.is_none() {
+            Some("stdin")
+        } else if stdout.is_none() {
+            Some("stdout")
+        } else if stderr.is_none() {
+            Some("stderr")
+        } else {
+            None
+        };
 
-        let stdin: SharedStdin = Arc::new(Mutex::new(Some(Box::new(stdin))));
+        let stdin: SharedStdin =
+            Arc::new(Mutex::new(stdin.map(|stdin| {
+                Box::new(stdin) as Box<dyn AsyncWrite + Send + Unpin>
+            })));
         let diagnostics = Arc::new(Mutex::new(BoundedDiagnostics::new(
             spec.name.clone(),
             limits.max_diagnostic_lines,
@@ -718,16 +731,32 @@ impl McpStdioClient {
         )));
         let list_changes = Arc::new(Mutex::new(BoundedListChanged::new(spec.name.clone())));
         let (tx, rx) = mpsc::channel(16);
-        let reader_task = spawn_stdout_reader(
-            spec.name.clone(),
-            stdout,
-            stdin.clone(),
-            tx,
-            limits.clone(),
-            redactor.clone(),
-            list_changes.clone(),
-        );
-        let stderr_task = spawn_stderr_reader(stderr, diagnostics.clone(), limits.clone());
+        // Do not return an error that discards an already spawned child. Queue
+        // setup failure before reader events so initialization cannot succeed.
+        if let Some(pipe) = missing_pipe {
+            tx.try_send(ReaderEvent::Error(McpClientError::new(
+                &spec.name,
+                McpPhase::Spawn,
+                McpErrorKind::Protocol(format!("child {pipe} was not piped")),
+            )))
+            .expect("new response channel has capacity");
+        }
+        let reader_task = match stdout {
+            Some(stdout) => spawn_stdout_reader(
+                spec.name.clone(),
+                stdout,
+                stdin.clone(),
+                tx,
+                limits.clone(),
+                redactor.clone(),
+                list_changes.clone(),
+            ),
+            None => tokio::spawn(async {}),
+        };
+        let stderr_task = match stderr {
+            Some(stderr) => spawn_stderr_reader(stderr, diagnostics.clone(), limits.clone()),
+            None => tokio::spawn(async {}),
+        };
 
         Ok(Self {
             server_name: spec.name,
@@ -746,6 +775,25 @@ impl McpStdioClient {
         })
     }
 
+    /// Perform the bounded handshake on an already retained client. Failure or
+    /// cancellation does not drop or shut down the client: its owner must do so.
+    /// Do not retry initialization on a partially initialized connection.
+    pub async fn initialize_with_timeout(&mut self) -> Result<(), McpClientError> {
+        let started = Instant::now();
+        match timeout(self.limits.startup_timeout, self.initialize()).await {
+            Ok(result) => result,
+            Err(_) => Err(McpClientError::new(
+                &self.server_name,
+                McpPhase::Initialize,
+                McpErrorKind::Timeout {
+                    operation: "startup".to_string(),
+                    elapsed: started.elapsed(),
+                },
+            )
+            .with_diagnostics(self.snapshot_diagnostics().await)),
+        }
+    }
+
     async fn initialize(&mut self) -> Result<(), McpClientError> {
         let result: InitializeResult = self
             .request(
@@ -1080,6 +1128,8 @@ impl McpStdioClient {
     }
 
     /// Close stdin and wait for process exit, falling back to terminate and kill.
+    /// Errors and cancellation retain the child for another attempt on this
+    /// client. Only confirmed reaping releases the process handle.
     pub async fn shutdown(&mut self) -> Result<ShutdownReport, McpClientError> {
         self.shutdown_started = true;
         let deadline = Instant::now() + self.limits.shutdown_timeout;
@@ -1091,64 +1141,55 @@ impl McpStdioClient {
         .await
         .is_ok();
 
-        let mut child = match self.child.take() {
-            Some(child) => child,
-            None => return Ok(ShutdownReport::already_finished()),
-        };
+        // Keep the actual process handle in the client across every await and
+        // error. Taking it before wait loses retry/reaping authority on cancel.
+        if self.child.is_none() {
+            return Ok(ShutdownReport::already_finished());
+        }
         if !closed {
             self.reader_task.abort();
-            return self.terminate_then_kill(child).await;
+            return self.terminate_then_kill().await;
         }
 
+        let child = self.child.as_mut().expect("checked child");
         if let Ok(Some(status)) = child.try_wait() {
-            self.reader_task.abort();
-            self.stderr_task.abort();
-            return Ok(ShutdownReport {
-                exit_status: Some(status),
-                terminated: false,
-                killed: false,
-            });
+            return Ok(self.finish_shutdown(status, false, false));
         }
 
         match timeout_at(deadline, child.wait()).await {
-            Ok(Ok(status)) => {
-                self.reader_task.abort();
-                self.stderr_task.abort();
-                Ok(ShutdownReport {
-                    exit_status: Some(status),
-                    terminated: false,
-                    killed: false,
-                })
-            }
+            Ok(Ok(status)) => Ok(self.finish_shutdown(status, false, false)),
             Ok(Err(err)) => Err(McpClientError::new(
                 &self.server_name,
                 McpPhase::Shutdown,
                 McpErrorKind::Io(self.redactor.redact(&err.to_string())),
             )
             .with_diagnostics(self.snapshot_diagnostics().await)),
-            Err(_) => self.terminate_then_kill(child).await,
+            Err(_) => self.terminate_then_kill().await,
         }
     }
 
-    async fn terminate_then_kill(
+    fn finish_shutdown(
         &mut self,
-        mut child: Child,
-    ) -> Result<ShutdownReport, McpClientError> {
-        let mut terminated = false;
-        let mut killed = false;
-        if send_terminate(&mut child).is_ok() {
-            terminated = true;
+        status: ExitStatus,
+        terminated: bool,
+        killed: bool,
+    ) -> ShutdownReport {
+        // No await separates confirmed reaping from releasing the handle.
+        drop(self.child.take());
+        self.reader_task.abort();
+        self.stderr_task.abort();
+        ShutdownReport {
+            exit_status: Some(status),
+            terminated,
+            killed,
         }
+    }
+
+    async fn terminate_then_kill(&mut self) -> Result<ShutdownReport, McpClientError> {
+        let child = self.child.as_mut().expect("shutdown retains child");
+        let terminated = send_terminate(child).is_ok();
         match timeout(self.limits.kill_timeout, child.wait()).await {
-            Ok(Ok(status)) => {
-                self.reader_task.abort();
-                self.stderr_task.abort();
-                Ok(ShutdownReport {
-                    exit_status: Some(status),
-                    terminated,
-                    killed,
-                })
-            }
+            Ok(Ok(status)) => Ok(self.finish_shutdown(status, terminated, false)),
             Ok(Err(err)) => Err(McpClientError::new(
                 &self.server_name,
                 McpPhase::Shutdown,
@@ -1163,7 +1204,6 @@ impl McpStdioClient {
                         McpErrorKind::Io(self.redactor.redact(&err.to_string())),
                     )
                 })?;
-                killed = true;
                 let status = timeout(self.limits.kill_timeout, child.wait())
                     .await
                     .map_err(|_| {
@@ -1183,13 +1223,7 @@ impl McpStdioClient {
                             McpErrorKind::Io(self.redactor.redact(&err.to_string())),
                         )
                     })?;
-                self.reader_task.abort();
-                self.stderr_task.abort();
-                Ok(ShutdownReport {
-                    exit_status: Some(status),
-                    terminated,
-                    killed,
-                })
+                Ok(self.finish_shutdown(status, terminated, true))
             }
         }
     }
@@ -1197,10 +1231,10 @@ impl McpStdioClient {
 
 impl Drop for McpStdioClient {
     fn drop(&mut self) {
-        if !self.shutdown_started {
-            if let Some(child) = &mut self.child {
-                let _ = child.start_kill();
-            }
+        // An attempted shutdown is not proof of exit. Errors and cancellation
+        // leave the child here so Drop must still make a best-effort kill.
+        if let Some(child) = &mut self.child {
+            let _ = child.start_kill();
         }
         self.reader_task.abort();
         self.stderr_task.abort();
@@ -1784,3 +1818,70 @@ fn send_terminate(child: &mut Child) -> Result<(), ()> {
 
 #[cfg(test)]
 mod deadline_tests;
+
+#[cfg(all(test, unix))]
+mod shutdown_retention_tests {
+    //! The real client and OS process are used; only the external server is a
+    //! local shell that acknowledges initialization and ignores TERM/EOF.
+    use super::*;
+    use std::future::{Future, poll_fn};
+    use std::task::Poll;
+
+    #[tokio::test]
+    async fn cancelled_graceful_or_terminate_wait_retries_reaping_the_same_child() {
+        for cancel_in_fallback in [false, true] {
+            let script = r#"trap '' TERM
+read init || exit 1
+printf '%s\n' '{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-11-25","capabilities":{},"serverInfo":{"name":"retained","version":"1"}}}'
+while :; do :; done
+"#;
+            let mut client = McpStdioClient::connect(
+                McpStdioServerSpec::new("retained", "/bin/sh").args(["-c", script]),
+                McpStdioLimits {
+                    shutdown_timeout: if cancel_in_fallback {
+                        Duration::ZERO
+                    } else {
+                        Duration::from_secs(30)
+                    },
+                    kill_timeout: Duration::from_secs(30),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("initialize after installing TERM handler");
+            let pid = client.child.as_ref().unwrap().id();
+            let mut shutdown = Box::pin(client.shutdown());
+            poll_fn(|cx| {
+                assert!(shutdown.as_mut().poll(cx).is_pending());
+                Poll::Ready(())
+            })
+            .await;
+            drop(shutdown);
+            assert_eq!(client.child.as_ref().and_then(Child::id), pid);
+            assert!(client.shutdown_started);
+
+            client.limits.shutdown_timeout = Duration::from_millis(20);
+            client.limits.kill_timeout = Duration::from_millis(50);
+            let report = timeout(Duration::from_secs(2), client.shutdown())
+                .await
+                .expect("retry bounded")
+                .expect("retry reaps original child");
+            assert!(
+                report.exit_status.is_some(),
+                "not an already-finished receipt"
+            );
+            assert!(report.killed, "TERM was ignored");
+            assert!(client.child.is_none());
+            let mut status = 0;
+            assert_eq!(
+                unsafe { libc::waitpid(pid.unwrap() as libc::pid_t, &mut status, libc::WNOHANG) },
+                -1
+            );
+            assert_eq!(
+                std::io::Error::last_os_error().raw_os_error(),
+                Some(libc::ECHILD),
+                "shutdown, not a later observer, reaped the original process"
+            );
+        }
+    }
+}

crates/runtime-api/src/lib.rs

@@ -86,6 +86,8 @@ pub enum WorkerRestoreState {
     Accepted,
     Rejected,
     RolledBack,
+    /// A later Stop superseded this restore; neither rollback nor prior Accepted is asserted.
+    Superseded,
     ReconciliationRequired,
 }
 
@@ -1578,13 +1580,24 @@ mod tests {
         )
         .unwrap();
         assert!(request.preparation.is_none());
-        for state in [
-            "accepted",
-            "rejected",
-            "rolled_back",
-            "reconciliation_required",
+        for (state, wire) in [
+            (WorkerRestoreState::Accepted, "accepted"),
+            (WorkerRestoreState::Rejected, "rejected"),
+            (WorkerRestoreState::RolledBack, "rolled_back"),
+            (WorkerRestoreState::Superseded, "superseded"),
+            (
+                WorkerRestoreState::ReconciliationRequired,
+                "reconciliation_required",
+            ),
         ] {
-            assert!(serde_json::from_value::<WorkerRestoreState>(serde_json::json!(state)).is_ok());
+            assert_eq!(
+                serde_json::from_value::<WorkerRestoreState>(serde_json::json!(wire)).unwrap(),
+                state
+            );
+            assert_eq!(
+                serde_json::to_value(state).unwrap(),
+                serde_json::json!(wire)
+            );
         }
         assert!(
             serde_json::from_value::<WorkerRestoreState>(serde_json::json!(

crates/server-api/src/lib.rs

@@ -552,8 +552,9 @@ pub struct WorkspaceWorkerSessionHistoryResponse {
 /// comparisons do not replace admission checks under the Runtime operation lock; conflicts do
 /// not trigger automatic token replacement.
 /// [`WorkerRestoreState`] Accepted means Restore committed, not that the subsequent run succeeded.
-/// Distinguish read-only preflight Rejected, confirmed RolledBack after side effects, and
-/// ReconciliationRequired needing settled-state checks. Response loss does not imply rollback.
+/// Distinguish read-only preflight Rejected, confirmed RolledBack after side effects,
+/// Superseded by a later Stop, and ReconciliationRequired needing settled-state checks.
+/// Response loss does not imply rollback.
 #[api(reqwest, axum, openapi)]
 pub trait ServerApi {
     #[get("/health", status = 200, error_status = 400)]
@@ -10579,8 +10580,8 @@ pub enum WorkerOperationState {
 /// Cross-layer outcome of an explicit Worker restore operation.
 ///
 /// `Rejected` is reserved for read-only preflight failures. Once live restore
-/// work starts, failure is either a confirmed `RolledBack` operation or a
-/// `ReconciliationRequired` result whose commit state must be reread/retried.
+/// work starts, distinguish confirmed `RolledBack`, unresolved
+/// `ReconciliationRequired`, and a later Stop's `Superseded` outcome.
 #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, JsonSchema)]
 #[cfg_attr(feature = "typescript", derive(ts_rs::TS))]
 #[serde(rename_all = "snake_case")]
@@ -10588,6 +10589,8 @@ pub enum WorkerRestoreState {
     Accepted,
     Rejected,
     RolledBack,
+    /// A later Stop superseded this restore; neither rollback nor prior Accepted is asserted.
+    Superseded,
     ReconciliationRequired,
 }
 
@@ -12681,6 +12684,7 @@ mod tests {
             (WorkerRestoreState::Accepted, "accepted"),
             (WorkerRestoreState::Rejected, "rejected"),
             (WorkerRestoreState::RolledBack, "rolled_back"),
+            (WorkerRestoreState::Superseded, "superseded"),
             (
                 WorkerRestoreState::ReconciliationRequired,
                 "reconciliation_required",

crates/session-store/src/worker_session_store.rs

@@ -194,6 +194,21 @@ impl WorkerSessionStore {
         Ok(store)
     }
 
+    /// Open a current retained Session without creating or migrating anything.
+    /// The live constraint is checked here and again by every later publication;
+    /// attaching the control store itself does not acquire publication authority.
+    pub fn open_existing_with_retention(
+        root: impl Into<PathBuf>,
+        retention: crate::retention::SessionRetentionStore,
+    ) -> Result<Self, StoreError> {
+        let mut store = Self::open_read_only(root)?;
+        if let Some(session) = *store.session_id.lock().unwrap() {
+            retention.ensure_live(session)?;
+        }
+        store.retention = Some(retention);
+        Ok(store)
+    }
+
     fn lock_publication(
         &self,
         session: SessionId,

crates/tui/src/backend_worker_picker.rs

@@ -785,6 +785,12 @@ mod tests {
     #[derive(Default)]
     struct PendingRuntimeBackend {
         operations: std::sync::Mutex<Vec<worker_runtime::execution::WorkerLifecycleOperationId>>,
+        owners: std::sync::Mutex<
+            std::collections::HashMap<
+                worker_runtime::execution::WorkerLifecycleOperationId,
+                worker_runtime::identity::WorkerRef,
+            >,
+        >,
     }
 
     impl worker_runtime::execution::WorkerExecutionBackend for PendingRuntimeBackend {
@@ -806,6 +812,21 @@ mod tests {
             )
         }
 
+        fn restore_recovery_disposition(
+            &self,
+            operation_id: worker_runtime::execution::WorkerLifecycleOperationId,
+            worker_ref: &worker_runtime::identity::WorkerRef,
+        ) -> Option<worker_runtime::execution::RestoreRecoveryDisposition> {
+            use worker_runtime::execution::RestoreRecoveryDisposition as D;
+            Some(
+                if self.owners.lock().unwrap().get(&operation_id) == Some(worker_ref) {
+                    D::Continue
+                } else {
+                    D::Absent
+                },
+            )
+        }
+
         fn restore_worker(
             &self,
             request: worker_runtime::execution::WorkerExecutionRestoreRequest,
@@ -813,6 +834,10 @@ mod tests {
             use worker_runtime::execution::*;
             let mut operations = self.operations.lock().unwrap();
             operations.push(request.operation_id);
+            self.owners
+                .lock()
+                .unwrap()
+                .insert(request.operation_id, request.worker_ref.clone());
             if operations.len() == 1 {
                 WorkerExecutionSpawnResult::ReconciliationRequired {
                     result: WorkerExecutionResult::errored(
@@ -1119,39 +1144,45 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn rejected_resume_does_not_observe_or_open_retained_state() {
-        let lifecycle = FakeBackendWorkerLifecycle {
-            restore_response: restore_response(BackendWorkerRestoreState::Rejected),
-            observation: session_observation(serde_json::json!({
-                "availability": "unavailable",
-                "reason": "retention_missing",
-                "message": "must not be opened"
-            })),
-            calls: std::sync::Mutex::new(Vec::new()),
-        };
-        let target = client::BackendRuntimeTarget::new(
-            "http://127.0.0.1:3000",
-            "workspace-a",
-            "runtime-a",
-            "worker-a",
-        );
+    async fn settled_nonaccepted_resume_does_not_retry_observe_or_open_retained_state() {
+        for state in [
+            BackendWorkerRestoreState::Rejected,
+            BackendWorkerRestoreState::RolledBack,
+            BackendWorkerRestoreState::Superseded,
+        ] {
+            let lifecycle = FakeBackendWorkerLifecycle {
+                restore_response: restore_response(state),
+                observation: session_observation(serde_json::json!({
+                    "availability": "unavailable",
+                    "reason": "retention_missing",
+                    "message": "must not be opened"
+                })),
+                calls: std::sync::Mutex::new(Vec::new()),
+            };
+            let target = client::BackendRuntimeTarget::new(
+                "http://127.0.0.1:3000",
+                "workspace-a",
+                "runtime-a",
+                "worker-a",
+            );
 
-        let error = prepare_selected_worker(
-            &lifecycle,
-            target,
-            BackendWorkerPickerIntent::Resume,
-            Some(client::BackendWorkerRestoreRequest {
-                expected_observation_token: "observed-worker-state".into(),
-                request_id: "resume-operation".into(),
-            }),
-            |_| Ok(false),
-        )
-        .await
-        .unwrap_err()
-        .to_string();
+            let error = prepare_selected_worker(
+                &lifecycle,
+                target,
+                BackendWorkerPickerIntent::Resume,
+                Some(client::BackendWorkerRestoreRequest {
+                    expected_observation_token: "observed-worker-state".into(),
+                    request_id: "resume-operation".into(),
+                }),
+                |_| panic!("settled nonaccepted restore must not offer recovery retry"),
+            )
+            .await
+            .unwrap_err()
+            .to_string();
 
-        assert!(error.contains("Rejected"));
-        assert_eq!(*lifecycle.calls.lock().unwrap(), ["restore"]);
+            assert!(error.contains(&format!("{state:?}")));
+            assert_eq!(*lifecycle.calls.lock().unwrap(), ["restore"]);
+        }
     }
 
     #[test]
@@ -1176,6 +1207,7 @@ mod tests {
         for state in [
             BackendWorkerRestoreState::Rejected,
             BackendWorkerRestoreState::RolledBack,
+            BackendWorkerRestoreState::Superseded,
             BackendWorkerRestoreState::ReconciliationRequired,
         ] {
             let mut response = restore_response(state);

crates/worker-runtime/Cargo.toml

@@ -19,15 +19,19 @@ required-features = ["ws-server", "fs-store"]
 [features]
 default = ["ws-server", "fs-store"]
 fs-store = ["dep:fs4"]
+# Offline fixture tooling only; never enabled by the product defaults.
+test-support = ["fs-store", "ws-server"]
+# Explicit process-test synchronization; independent of offline fixture tooling.
+restore-checkpoints = ["fs-store"]
 http-server = ["dep:axum", "dep:tower", "dep:reqwest"]
-ws-server = ["http-server", "axum/ws", "dep:futures", "tokio/sync"]
+ws-server = ["http-server", "axum/ws", "tokio/sync"]
 
 [dependencies]
 agen.workspace = true
 async-trait.workspace = true
 base64.workspace = true
 axum = { workspace = true, optional = true }
-futures = { workspace = true, optional = true }
+futures.workspace = true
 decodal.workspace = true
 flow = { path = "../flow" }
 fs4 = { workspace = true, optional = true }

crates/worker-runtime/README.md

@@ -8,3 +8,61 @@ Manages Worker creation, stopping, and Restore, execution connections within the
 - [`execution`] — execution backends and operation-specific results.
 - [`working_directory`] — repository materialization, session resources, and safe deletion.
 - [`workspace_issuer`] — Workspace signature verification and replay prevention.
+
+## Offline stopped-Worker fixtures
+
+The nondefault `test-support` feature exposes `test_support::StoppedWorkerTemplate`.
+It is intentionally limited to current-schema, stopped, attachment-free artificial
+conversations with one Segment, text history, and no pending execution/activation.
+It does not add a product endpoint or change create/restore authorization.
+
+1. With the source Runtime offline, call `StoppedWorkerTemplate::extract(worker_dir)`.
+   The source must have its existing Runtime owner lock and identity. Extraction is
+   read-only; the returned serde template excludes source auth, saved Manifest,
+   request/config references, run directories, and Runtime operation history.
+2. Initialize **new** Runtime identity and Workspace trust in a fresh destination
+   using normal setup. Preserve the template's Runtime and Workspace IDs.
+3. Call `template.install(&options, request, &resolved_manifest)` before starting
+   the Runtime. Supply a fresh typed `CreateWorkerRequest` (including config/profile
+   references and trusted memory settings) and `manifest::WorkerManifest` resolved
+   for that environment. Its Worker name must equal `template.worker_name()`.
+4. Start normally and use the ordinary explicit restore operation. Installation
+   itself creates no execution and leaves the Worker stopped with a fresh guard.
+
+The template preserves the original Session manifest/JSONL bytes, including stable
+entry IDs, timestamps, and settled submission receipts. It rejects resumable queues
+and Feature invocation history. A saved system prompt is required for restore;
+checkpoint receipts must match the Worker's typed receipt schema. Preserved records
+must use canonical writer field shapes: unknown/duplicate fields (including nested
+fields), discarded optional fields, and opaque Feature history payloads are rejected,
+not stripped or rewritten. Portability/secret-marker checks do not replace
+review of the supplied artificial conversation or its independently verified hashes.
+The installer refuses any destination content other than initialized `auth.toml`;
+interrupted publication is fail-closed, not an overwrite/recovery facility.
+
+## Owned-process Restore checkpoints
+
+The independent, nondefault `restore-checkpoints` feature exists only for explicit
+process tests in the shared `tests/e2e` harness. Offline `test-support` does not enable
+it. Normal binaries reject `--test-restore-checkpoint`; no environment variable or
+HTTP endpoint activates this facility. Do not deploy a checkpoint-enabled binary
+as a replacement for the default product.
+
+A test-enabled Runtime accepts `--test-restore-checkpoint <config.json>` with a
+bounded regular JSON file selecting a loopback socket, Worker ID, one stage and a
+diagnostic nonce. The nonce is only a handshake label, never execution identity,
+request authority or routing state. The one-shot stages are real prepared candidate
+with `NotCommitted`, and durable `Committed` before activation/terminal receipt.
+Before acknowledging, the existing bounded canonical reader validates the exact
+Worker/operation/decision, syncs its file/directories and requires no saved receipt.
+No pending JSON is fabricated and no lifecycle work is skipped.
+
+The Runtime sends bounded newline-delimited JSON evidence, then waits at most eight
+seconds for byte `R`. EOF, invalid reply, unavailable peer or deadline releases the
+wait into the original operation. The Runtime-state mutex is not held across that
+wait; existing per-Worker ownership still covers the actual candidate and cleanup.
+The harness verifies the ack, kills/reaps only its owned child and inspects the
+canonical aggregate offline before restarting the same roots/identity without the
+checkpoint flag or another fixture install. These tests are a separate cfg/binary
+lane from featureless product regression; they do not certify arbitrary OS failure,
+external power loss or a real model call.

crates/worker-runtime/src/execution.rs

@@ -215,6 +215,38 @@ impl WorkerExecutionResult {
     }
 }
 
+// Private one-shot synchronization at real publication boundaries. Disconnect
+// releases the waiter during unwinding; timeout bounds a forgotten test owner.
+#[cfg(test)]
+#[derive(Debug)]
+pub(crate) struct RestoreTestGate<T> {
+    pub(crate) entered: std::sync::mpsc::Sender<T>,
+    pub(crate) release: std::sync::mpsc::Receiver<()>,
+}
+
+#[cfg(test)]
+impl<T> RestoreTestGate<T> {
+    pub(crate) fn wait(self, evidence: T) {
+        let _ = self.entered.send(evidence);
+        assert!(
+            !matches!(
+                self.release
+                    .recv_timeout(std::time::Duration::from_secs(30)),
+                Err(std::sync::mpsc::RecvTimeoutError::Timeout)
+            ),
+            "publication test gate timed out"
+        );
+    }
+}
+
+#[cfg(all(test, feature = "ws-server"))]
+#[derive(Debug)]
+pub(crate) struct RestoreFlushFault {
+    pub(crate) remaining_successes: usize,
+    pub(crate) gate: RestoreTestGate<protocol::Event>,
+    pub(crate) restored: std::sync::mpsc::Sender<Vec<protocol::Event>>,
+}
+
 /// Runtime hooks available to an execution backend for one Worker.
 #[cfg(feature = "ws-server")]
 type WorkerObservationPublisher = Arc<
@@ -242,10 +274,15 @@ enum WorkerExecutionPublicationState {
 #[derive(Clone)]
 pub struct WorkerExecutionContext {
     worker_ref: WorkerRef,
+    restore_progress: Option<Arc<dyn Fn(WorkerLifecycleOperationId) + Send + Sync>>,
+    restore_repair_authorizer:
+        Option<Arc<dyn Fn(WorkerLifecycleOperationId) -> Result<(), RuntimeError> + Send + Sync>>,
     #[cfg(feature = "ws-server")]
     observation_publisher: WorkerObservationPublisher,
     #[cfg(feature = "ws-server")]
     publication_state: Arc<Mutex<WorkerExecutionPublicationState>>,
+    #[cfg(all(test, feature = "ws-server"))]
+    flush_fault: Arc<Mutex<Option<RestoreFlushFault>>>,
 }
 
 impl WorkerExecutionContext {
@@ -256,7 +293,11 @@ impl WorkerExecutionContext {
     ) -> Self {
         Self {
             worker_ref,
+            restore_progress: None,
+            restore_repair_authorizer: None,
             observation_publisher,
+            #[cfg(test)]
+            flush_fault: Arc::default(),
             publication_state: Arc::new(Mutex::new(WorkerExecutionPublicationState::Active)),
         }
     }
@@ -268,7 +309,11 @@ impl WorkerExecutionContext {
     ) -> Self {
         Self {
             worker_ref,
+            restore_progress: None,
+            restore_repair_authorizer: None,
             observation_publisher,
+            #[cfg(test)]
+            flush_fault: Arc::default(),
             publication_state: Arc::new(Mutex::new(WorkerExecutionPublicationState::Pending(
                 VecDeque::new(),
             ))),
@@ -277,12 +322,56 @@ impl WorkerExecutionContext {
 
     #[cfg(not(feature = "ws-server"))]
     pub(crate) fn new(worker_ref: WorkerRef) -> Self {
-        Self { worker_ref }
+        Self {
+            worker_ref,
+            restore_progress: None,
+            restore_repair_authorizer: None,
+        }
     }
 
     #[cfg(not(feature = "ws-server"))]
     pub(crate) fn candidate(worker_ref: WorkerRef) -> Self {
-        Self { worker_ref }
+        Self {
+            worker_ref,
+            restore_progress: None,
+            restore_repair_authorizer: None,
+        }
+    }
+
+    pub(crate) fn with_restore_repair_authorizer(
+        mut self,
+        authorizer: Arc<
+            dyn Fn(WorkerLifecycleOperationId) -> Result<(), RuntimeError> + Send + Sync,
+        >,
+    ) -> Self {
+        self.restore_repair_authorizer = Some(authorizer);
+        self
+    }
+
+    pub(crate) fn authorize_restore_source_repair(
+        &self,
+        operation: WorkerLifecycleOperationId,
+    ) -> Result<(), RuntimeError> {
+        match &self.restore_repair_authorizer {
+            Some(authorizer) => authorizer(operation),
+            // Direct factory/component users own their source repair separately;
+            // only Runtime contexts carry a Runtime aggregate authority callback.
+            None => Ok(()),
+        }
+    }
+
+    pub(crate) fn with_restore_progress(
+        mut self,
+        progress: Arc<dyn Fn(WorkerLifecycleOperationId) + Send + Sync>,
+    ) -> Self {
+        self.restore_progress = Some(progress);
+        self
+    }
+
+    pub(crate) fn restore_progressed(&self, operation_id: WorkerLifecycleOperationId) {
+        if let Some(progress) = &self.restore_progress {
+            progress(operation_id);
+        }
     }
 
     pub fn worker_ref(&self) -> &WorkerRef {
@@ -335,20 +424,65 @@ impl WorkerExecutionContext {
             let Some(event) = event else {
                 return Ok(());
             };
-            if let Err(error) = (self.observation_publisher)(self.worker_ref.clone(), event.clone())
-            {
+            #[cfg(test)]
+            let fault = {
+                let mut fault = self.flush_fault.lock().unwrap();
+                if fault.as_ref().is_some_and(|f| f.remaining_successes == 0) {
+                    fault.take()
+                } else {
+                    if let Some(fault) = fault.as_mut() {
+                        fault.remaining_successes -= 1;
+                    }
+                    None
+                }
+            };
+            #[cfg(test)]
+            let mut restored = None;
+            #[cfg(test)]
+            let publication = if let Some(fault) = fault {
+                restored = Some(fault.restored);
+                fault.gate.wait(event.clone());
+                Err(RuntimeError::InvalidRequest(
+                    "injected buffered publication failure".into(),
+                ))
+            } else {
+                (self.observation_publisher)(self.worker_ref.clone(), event.clone())
+            };
+            #[cfg(not(test))]
+            let publication = (self.observation_publisher)(self.worker_ref.clone(), event.clone());
+            if let Err(error) = publication {
                 let mut state = self
                     .publication_state
                     .lock()
                     .map_err(|_| RuntimeError::StatePoisoned)?;
                 if let WorkerExecutionPublicationState::Activating(events) = &mut *state {
                     events.push_front(event);
+                    #[cfg(test)]
+                    if let Some(restored) = restored {
+                        let _ = restored.send(events.iter().cloned().collect());
+                    }
                 }
                 return Err(error);
             }
         }
     }
 
+    #[cfg(all(test, feature = "ws-server"))]
+    pub(crate) fn test_fail_flush(&self, fault: RestoreFlushFault) {
+        assert!(self.flush_fault.lock().unwrap().replace(fault).is_none());
+    }
+
+    #[cfg(all(test, feature = "ws-server"))]
+    pub(crate) fn test_buffered_events(&self) -> Vec<protocol::Event> {
+        match &*self.publication_state.lock().unwrap() {
+            WorkerExecutionPublicationState::Pending(events)
+            | WorkerExecutionPublicationState::Activating(events) => {
+                events.iter().cloned().collect()
+            }
+            _ => Vec::new(),
+        }
+    }
+
     #[cfg(not(feature = "ws-server"))]
     pub(crate) fn activate_candidate(&self) -> Result<(), RuntimeError> {
         Ok(())
@@ -424,6 +558,7 @@ pub struct WorkerExecutionSpawnRequest {
 /// Request passed to a [`WorkerExecutionBackend`] when restoring a persisted Worker.
 #[derive(Clone, Debug)]
 pub struct WorkerExecutionRestoreRequest {
+    pub source_repair_possible: bool,
     pub operation_id: WorkerLifecycleOperationId,
     pub worker_ref: WorkerRef,
     pub request: crate::catalog::CreateWorkerRequest,
@@ -565,6 +700,19 @@ impl fmt::Debug for WorkerProtocolTransport {
     }
 }
 
+/// Derived resource disposition for an existing restore owner. Not durable
+/// authority: Runtime must reread its decision before choosing any action.
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum RestoreRecoveryDisposition {
+    /// No live or pending owner for this exact operation. Not cleanup evidence.
+    Absent,
+    InFlight,
+    Continue,
+    Discard,
+    /// A retained cleanup-only owner; never reconstruct or activate it.
+    Reclaim,
+}
+
 /// The connection selects this backend; WorkerRef identifies an execution,
 /// never a capability or authorization token.
 pub trait WorkerExecutionBackend: Send + Sync + 'static {
@@ -633,6 +781,33 @@ pub trait WorkerExecutionBackend: Send + Sync + 'static {
         Ok(())
     }
 
+    /// Completed resource-owner evidence retained while Runtime settlement may
+    /// still need retry. This is not a new request or execution authorization.
+    fn completed_restore_failure(
+        &self,
+        _operation_id: WorkerLifecycleOperationId,
+        _worker_ref: &WorkerRef,
+    ) -> Option<WorkerExecutionSpawnResult> {
+        None
+    }
+
+    /// Opt in to Runtime-owned autonomous restore continuation. Pending tasks
+    /// signal completion through the original execution context; callbacks must
+    /// not wait for Runtime locks. Independent of resource disposition support.
+    fn supports_restore_continuation(&self) -> bool {
+        false
+    }
+
+    /// Describe exact operation resource ownership for warm/cold recovery
+    /// dispatch. This does not opt in to callbacks or autonomous continuation.
+    fn restore_recovery_disposition(
+        &self,
+        _operation_id: WorkerLifecycleOperationId,
+        _worker_ref: &WorkerRef,
+    ) -> Option<RestoreRecoveryDisposition> {
+        None
+    }
+
     fn reconcile_restore(
         &self,
         request: WorkerExecutionRestoreRequest,
@@ -640,6 +815,42 @@ pub trait WorkerExecutionBackend: Send + Sync + 'static {
         self.restore_worker(request)
     }
 
+    /// Reclaim an uncommitted operation after restart without preparing,
+    /// repairing, or activating a replacement candidate. Runtime must establish
+    /// the durable uncommitted decision before calling this seam. Unsupported
+    /// cleanup is uncertainty, never a successful rollback.
+    fn reclaim_uncommitted_restore(
+        &self,
+        _request: WorkerExecutionRestoreRequest,
+    ) -> WorkerExecutionSpawnResult {
+        WorkerExecutionSpawnResult::ReconciliationRequired {
+            result: WorkerExecutionResult::unsupported(
+                WorkerExecutionOperation::Restore,
+                "backend cannot reclaim an uncommitted restore after restart",
+            ),
+            worker_state: None,
+            workdir_attachments: Vec::new(),
+        }
+    }
+
+    /// Discard only the exact unpublished operation after Runtime has proved
+    /// its durable decision uncommitted. This is not Stop and must not affect a
+    /// connected execution. RolledBack requires confirmed cleanup and no repair.
+    fn discard_uncommitted_restore(
+        &self,
+        _operation_id: WorkerLifecycleOperationId,
+        _worker_ref: &WorkerRef,
+    ) -> WorkerExecutionSpawnResult {
+        WorkerExecutionSpawnResult::ReconciliationRequired {
+            result: WorkerExecutionResult::unsupported(
+                WorkerExecutionOperation::Restore,
+                "backend cannot confirm unpublished restore cleanup",
+            ),
+            worker_state: None,
+            workdir_attachments: Vec::new(),
+        }
+    }
+
     fn activate_restored_worker(
         &self,
         _operation_id: WorkerLifecycleOperationId,
@@ -901,6 +1112,28 @@ impl WorkerExecutionBackendRef {
         self.backend.preflight_restore(request)
     }
 
+    pub(crate) fn completed_restore_failure(
+        &self,
+        operation_id: WorkerLifecycleOperationId,
+        worker_ref: &WorkerRef,
+    ) -> Option<WorkerExecutionSpawnResult> {
+        self.backend
+            .completed_restore_failure(operation_id, worker_ref)
+    }
+
+    pub(crate) fn supports_restore_continuation(&self) -> bool {
+        self.backend.supports_restore_continuation()
+    }
+
+    pub(crate) fn restore_recovery_disposition(
+        &self,
+        operation_id: WorkerLifecycleOperationId,
+        worker_ref: &WorkerRef,
+    ) -> Option<RestoreRecoveryDisposition> {
+        self.backend
+            .restore_recovery_disposition(operation_id, worker_ref)
+    }
+
     pub(crate) fn reconcile_restore(
         &self,
         request: WorkerExecutionRestoreRequest,
@@ -908,6 +1141,22 @@ impl WorkerExecutionBackendRef {
         self.backend.reconcile_restore(request)
     }
 
+    pub(crate) fn reclaim_uncommitted_restore(
+        &self,
+        request: WorkerExecutionRestoreRequest,
+    ) -> WorkerExecutionSpawnResult {
+        self.backend.reclaim_uncommitted_restore(request)
+    }
+
+    pub(crate) fn discard_uncommitted_restore(
+        &self,
+        operation_id: WorkerLifecycleOperationId,
+        worker_ref: &WorkerRef,
+    ) -> WorkerExecutionSpawnResult {
+        self.backend
+            .discard_uncommitted_restore(operation_id, worker_ref)
+    }
+
     pub(crate) fn activate_restored_worker(
         &self,
         operation_id: WorkerLifecycleOperationId,

crates/worker-runtime/src/http_server.rs

@@ -25,6 +25,7 @@ use crate::interaction::{WorkerInput, WorkerInteractionAck};
 use crate::management::{RuntimeSummary, WorkerDeleteResult};
 #[cfg(feature = "ws-server")]
 use crate::observation::WorkerObservationCursor;
+#[cfg(feature = "fs-store")]
 use crate::retention::{
     WorkerRetentionExecutionRequest, WorkerRetentionExecutionResult, WorkerRetentionInventory,
 };
@@ -1568,6 +1569,7 @@ fn protocol_error_event(message: impl Into<String>) -> protocol::Event {
     }
 }
 
+#[cfg(feature = "fs-store")]
 #[allow(dead_code)]
 async fn worker_retention_inventory(
     State(state): State<RuntimeHttpState>,
@@ -1589,6 +1591,7 @@ async fn worker_retention_inventory(
         .map_err(RuntimeHttpRestError::runtime)
 }
 
+#[cfg(feature = "fs-store")]
 #[allow(dead_code)]
 async fn execute_worker_retention(
     State(state): State<RuntimeHttpState>,
@@ -2654,6 +2657,87 @@ mod tests {
         .unwrap()
     }
 
+    // Generated management routes and real Workspace capability middleware stay
+    // connected even when this process has no filesystem retention provider.
+    #[cfg(not(feature = "fs-store"))]
+    #[tokio::test]
+    async fn memory_retention_routes_preserve_auth_and_return_typed_unavailable() {
+        let identity = RuntimeIdentityMaterial::generate("workspace-key").unwrap();
+        let trust = workspace_trust(&identity);
+        let runtime = Runtime::new_memory();
+        let app = runtime_http_router_with_workspace_auth(
+            runtime,
+            None,
+            workspace_test_auth(vec![trust.clone()]),
+        );
+        let worker_id = WorkerId::now_v7();
+        let body = runtime_api::WorkerRetentionExecutionRequest {
+            operation_id: "memory-retention".into(),
+            input_fingerprint: "input".into(),
+            archive_id: None,
+            workspace_id: "workspace-a".into(),
+            source_runtime_id: "runtime-test".into(),
+            worker_id,
+            source_created_at: "2026-01-01T00:00:00Z".into(),
+            removed_at: "2026-01-02T00:00:00Z".into(),
+            effective_profile: None,
+            retention_class: None,
+            policy_id: "policy".into(),
+            session_disposition: runtime_api::SessionDisposition::Purge,
+            diagnostics_disposition: runtime_api::DiagnosticsDisposition::Purge,
+        };
+        for (method, path, body) in [
+            (
+                "GET",
+                format!("/v1/workers/{worker_id}/retention/inventory"),
+                vec![],
+            ),
+            (
+                "POST",
+                format!("/v1/workers/{worker_id}/retention/execute"),
+                serde_json::to_vec(&body).unwrap(),
+            ),
+        ] {
+            let unsigned = app
+                .clone()
+                .oneshot(
+                    Request::builder()
+                        .method(method)
+                        .uri(&path)
+                        .header(header::CONTENT_TYPE, "application/json")
+                        .body(Body::from(body.clone()))
+                        .unwrap(),
+                )
+                .await
+                .unwrap();
+            assert_eq!(unsigned.status(), StatusCode::UNAUTHORIZED);
+            let claims = signed_claims(&trust, method, &path, &body);
+            let response =
+                signed_request(app.clone(), &identity, &claims, body, "workspace-a").await;
+            assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
+            let error: RuntimeHttpErrorResponse = read_json(response).await;
+            assert_eq!(error.error.code, "worker_retention_unavailable");
+        }
+        for (mut value, expected_status, expected_code) in [
+            (body.clone(), StatusCode::NOT_FOUND, "worker_not_found"),
+            (body.clone(), StatusCode::BAD_REQUEST, "worker_id_mismatch"),
+        ] {
+            if expected_status == StatusCode::NOT_FOUND {
+                value.workspace_id = "workspace-b".into();
+            } else {
+                value.worker_id = WorkerId::now_v7();
+            }
+            let path = format!("/v1/workers/{worker_id}/retention/execute");
+            let bytes = serde_json::to_vec(&value).unwrap();
+            let claims = signed_claims(&trust, "POST", &path, &bytes);
+            let response =
+                signed_request(app.clone(), &identity, &claims, bytes, "workspace-a").await;
+            assert_eq!(response.status(), expected_status);
+            let error: RuntimeHttpErrorResponse = read_json(response).await;
+            assert_eq!(error.error.code, expected_code);
+        }
+    }
+
     #[tokio::test]
     async fn ordinary_signed_ping_succeeds_without_handshake_and_rejects_replay() {
         let identity = RuntimeIdentityMaterial::generate("workspace-key").unwrap();

crates/worker-runtime/src/http_server/runtime_management_api.rs

@@ -719,6 +719,9 @@ impl runtime_api::RuntimeApi for RuntimeManagementApi {
                         server_api::WorkerRestoreState::RolledBack => {
                             runtime_api::WorkerRestoreState::RolledBack
                         }
+                        server_api::WorkerRestoreState::Superseded => {
+                            runtime_api::WorkerRestoreState::Superseded
+                        }
                         server_api::WorkerRestoreState::ReconciliationRequired => {
                             runtime_api::WorkerRestoreState::ReconciliationRequired
                         }
@@ -769,6 +772,9 @@ impl runtime_api::RuntimeApi for RuntimeManagementApi {
                 server_api::WorkerRestoreState::RolledBack => {
                     runtime_api::WorkerRestoreState::RolledBack
                 }
+                server_api::WorkerRestoreState::Superseded => {
+                    runtime_api::WorkerRestoreState::Superseded
+                }
                 server_api::WorkerRestoreState::ReconciliationRequired => {
                     runtime_api::WorkerRestoreState::ReconciliationRequired
                 }
@@ -881,13 +887,25 @@ impl runtime_api::RuntimeApi for RuntimeManagementApi {
                 )
             })?;
         let worker_ref = worker_ref_for(&self.state.runtime, worker_id).map_err(api_error)?;
-        let inventory = self
-            .state
-            .runtime
-            .worker_retention_inventory(&scope.workspace_id, &worker_ref)
-            .map_err(RuntimeHttpRestError::runtime)
-            .map_err(api_error)?;
-        response(inventory)
+        #[cfg(feature = "fs-store")]
+        {
+            let inventory = self
+                .state
+                .runtime
+                .worker_retention_inventory(&scope.workspace_id, &worker_ref)
+                .map_err(RuntimeHttpRestError::runtime)
+                .map_err(api_error)?;
+            response(inventory)
+        }
+        #[cfg(not(feature = "fs-store"))]
+        {
+            let _ = (scope, worker_ref);
+            Err(runtime_api::RuntimeApiError::new(
+                503,
+                "worker_retention_unavailable",
+                "Worker retention requires filesystem storage",
+            ))
+        }
     }
 
     async fn execute_retention(
@@ -905,27 +923,36 @@ impl runtime_api::RuntimeApi for RuntimeManagementApi {
                 )
             })?;
         let worker_ref = worker_ref_for(&self.state.runtime, worker_id).map_err(api_error)?;
-        let request: WorkerRetentionExecutionRequest = request(value)?;
-        if request.workspace_id != scope.workspace_id {
+        if value.workspace_id != scope.workspace_id {
             return Err(runtime_api::RuntimeApiError::new(
                 StatusCode::NOT_FOUND.as_u16(),
                 "worker_not_found",
                 "worker does not belong to the authenticated Workspace",
             ));
         }
-        if request.worker_id != worker_ref.worker_id {
+        if value.worker_id != worker_ref.worker_id {
             return Err(runtime_api::RuntimeApiError::new(
                 StatusCode::BAD_REQUEST.as_u16(),
                 "worker_id_mismatch",
                 "Worker retention request worker_id does not match route worker_id",
             ));
         }
-        let result = self
-            .state
-            .runtime
-            .execute_worker_retention(&request)
-            .map_err(RuntimeHttpRestError::runtime)
-            .map_err(api_error)?;
-        response(result)
+        #[cfg(feature = "fs-store")]
+        {
+            let request: WorkerRetentionExecutionRequest = request(value)?;
+            let result = self
+                .state
+                .runtime
+                .execute_worker_retention(&request)
+                .map_err(RuntimeHttpRestError::runtime)
+                .map_err(api_error)?;
+            response(result)
+        }
+        #[cfg(not(feature = "fs-store"))]
+        Err(runtime_api::RuntimeApiError::new(
+            503,
+            "worker_retention_unavailable",
+            "Worker retention requires filesystem storage",
+        ))
     }
 }

crates/worker-runtime/src/lib.rs

@@ -16,6 +16,11 @@ pub mod management;
 pub mod observation;
 pub mod profile_archive;
 pub mod resource;
+#[cfg(feature = "restore-checkpoints")]
+pub mod restore_checkpoints;
+#[cfg(test)]
+#[path = "../../../tests/support/restore_environment.rs"]
+mod restore_test_environment;
 #[cfg(feature = "fs-store")]
 pub mod retention;
 mod runtime;
@@ -32,6 +37,8 @@ pub mod workspace_issuer;
 #[cfg(feature = "http-server")]
 pub mod workspace_request;
 
+#[cfg(feature = "test-support")]
+pub use fs_store::test_support;
 #[cfg(feature = "fs-store")]
 pub use fs_store::{FsRuntimeStore, FsRuntimeStoreOptions};
 pub use management::RuntimeOptions;

crates/worker-runtime/src/main.rs

@@ -117,6 +117,12 @@ fn run() -> Result<(), ProcessError> {
         let listener = tokio::net::TcpListener::bind(config.http.bind_addr).await?;
         let local_addr = listener.local_addr()?;
         let worker_runtime = build_runtime(&config)?;
+        #[cfg(feature = "restore-checkpoints")]
+        if let Some(checkpoint) = &config.test_restore_checkpoint {
+            worker_runtime
+                .install_test_restore_checkpoint(checkpoint.clone())
+                .map_err(ProcessError::Runtime)?;
+        }
         eprintln!(
             "yoi-runtime listening on {local_addr}; intended client is a trusted backend/proxy, not a browser"
         );
@@ -391,6 +397,21 @@ where
         let (flag, inline_value) = split_flag_value(arg)?;
         match flag.as_str() {
             "--help" | "-h" => return Ok(None),
+            #[cfg(feature = "restore-checkpoints")]
+            "--test-restore-checkpoint" => {
+                let path = take_value(&flag, inline_value, &mut args)?;
+                if config.test_restore_checkpoint.is_some() {
+                    return Err(ProcessError::usage(
+                        "--test-restore-checkpoint may only be specified once".into(),
+                    ));
+                }
+                config.test_restore_checkpoint = Some(
+                    worker_runtime::restore_checkpoints::RestoreCheckpointConfig::from_path(
+                        Path::new(&path),
+                    )
+                    .map_err(ProcessError::Runtime)?,
+                );
+            }
             "--bind" => {
                 let value = take_value(&flag, inline_value, &mut args)?;
                 config.http.bind_addr = value.parse::<SocketAddr>().map_err(|error| {
@@ -465,6 +486,12 @@ where
     }
 
     apply_store_selection(&mut config);
+    #[cfg(feature = "restore-checkpoints")]
+    if config.test_restore_checkpoint.is_some() && config.no_store {
+        return Err(ProcessError::usage(
+            "--test-restore-checkpoint requires an fs Runtime store".into(),
+        ));
+    }
     Ok(Some(config))
 }
 
@@ -521,6 +548,8 @@ struct ProcessConfig {
     no_store: bool,
     backend_resource_endpoint: Option<String>,
     backend_resource_token: Option<String>,
+    #[cfg(feature = "restore-checkpoints")]
+    test_restore_checkpoint: Option<worker_runtime::restore_checkpoints::RestoreCheckpointConfig>,
 }
 
 #[derive(Clone, Debug, PartialEq, Eq)]
@@ -542,6 +571,8 @@ impl ProcessConfig {
             no_store: false,
             backend_resource_endpoint: None,
             backend_resource_token: None,
+            #[cfg(feature = "restore-checkpoints")]
+            test_restore_checkpoint: None,
         })
     }
 
@@ -1227,6 +1258,47 @@ Auth commands:
 mod tests {
     use super::*;
 
+    #[test]
+    fn checkpoint_cli_is_absent_by_default() {
+        let config = parse_args([] as [&str; 0]).unwrap().unwrap();
+        #[cfg(feature = "restore-checkpoints")]
+        assert!(config.test_restore_checkpoint.is_none());
+        #[cfg(not(feature = "restore-checkpoints"))]
+        {
+            let _ = config;
+            let error = parse_args(["--test-restore-checkpoint", "never-opened.json"]).unwrap_err();
+            assert!(error.to_string().contains("unknown argument"));
+        }
+    }
+
+    #[cfg(feature = "restore-checkpoints")]
+    #[test]
+    fn checkpoint_cli_requires_valid_explicit_file_and_fs_store() {
+        use worker_runtime::restore_checkpoints::{
+            RestoreCheckpointConfig, RestoreCheckpointStage,
+        };
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("checkpoint.json");
+        let checkpoint = RestoreCheckpointConfig {
+            endpoint: "127.0.0.1:1234".parse().unwrap(),
+            worker_id: protocol::WorkerId::now_v7(),
+            stage: RestoreCheckpointStage::CommittedPreterminal,
+            nonce: "cli-test".into(),
+        };
+        std::fs::write(&path, serde_json::to_vec(&checkpoint).unwrap()).unwrap();
+        let args = ["--test-restore-checkpoint", path.to_str().unwrap()];
+        let config = parse_args(args).unwrap().unwrap();
+        assert_eq!(config.test_restore_checkpoint, Some(checkpoint));
+        assert!(parse_args([args[0], args[1], "--no-store"]).is_err());
+        assert!(parse_args([args[0], args[1], args[0], args[1]]).is_err());
+        assert!(parse_args([args[0]]).is_err());
+        std::fs::write(&path, b"private-invalid-body").unwrap();
+        let error = parse_args(args).unwrap_err().to_string();
+        assert!(!error.contains("private-invalid-body"));
+        std::fs::remove_file(&path).unwrap();
+        assert!(parse_args(args).is_err());
+    }
+
     #[test]
     fn help_does_not_start_server() {
         assert!(parse_args(["--help"]).unwrap().is_none());

crates/worker-runtime/src/resource.rs

@@ -1,6 +1,8 @@
+#[cfg(feature = "http-server")]
 use crate::auth::BACKEND_RESOURCE_FETCH_PERMISSION;
 use crate::identity::WorkerId;
 use crate::profile_archive::{ProfileSourceArchive, ProfileSourceArchiveRef, sha256_hex};
+#[cfg(feature = "http-server")]
 use crate::workspace_request::RuntimeWorkspaceRequestClient;
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};

crates/worker-runtime/src/restore_checkpoints.rs

@@ -0,0 +1,144 @@
+//! Explicit, ephemeral synchronization for owned Runtime process tests only.
+//!
+//! The nonce authenticates no Runtime action and is never a routing identity.
+//! Socket failure, EOF and the fixed deadline only release the gate; the lifecycle
+//! owner continues its original operation and remains responsible for cleanup.
+
+use std::fs::OpenOptions;
+use std::io::{Read, Write};
+use std::net::{SocketAddr, TcpStream};
+use std::path::Path;
+use std::time::{Duration, Instant};
+
+use protocol::WorkerId;
+use serde::{Deserialize, Serialize};
+
+use crate::error::RuntimeError;
+use crate::execution::WorkerLifecycleOperationId;
+
+pub const RESTORE_CHECKPOINT_TIMEOUT: Duration = Duration::from_secs(8);
+pub const MAX_RESTORE_CHECKPOINT_CONFIG_BYTES: usize = 4096;
+pub const MAX_RESTORE_CHECKPOINT_ACK_BYTES: usize = 4096;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum RestoreCheckpointStage {
+    PreparedNotCommitted,
+    CommittedPreterminal,
+    ActivatedPreterminal,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum RestoreCheckpointCommit {
+    NotCommitted,
+    Committed,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RestoreCheckpointConfig {
+    pub endpoint: SocketAddr,
+    pub worker_id: WorkerId,
+    pub stage: RestoreCheckpointStage,
+    pub nonce: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct RestoreCheckpointAck {
+    pub nonce: String,
+    pub stage: RestoreCheckpointStage,
+    pub runtime_id: String,
+    pub worker_id: WorkerId,
+    pub operation_id: WorkerLifecycleOperationId,
+    pub commit: RestoreCheckpointCommit,
+    pub receipt_absent: bool,
+    pub pid: u32,
+}
+
+fn invalid_config() -> RuntimeError {
+    // Do not include serde errors, paths or input contents in process diagnostics.
+    RuntimeError::InvalidRequest("invalid test restore checkpoint configuration".into())
+}
+
+impl RestoreCheckpointConfig {
+    pub fn validate(&self) -> Result<(), RuntimeError> {
+        if !self.endpoint.ip().is_loopback()
+            || self.endpoint.port() == 0
+            || self.nonce.is_empty()
+            || self.nonce.len() > 256
+            || self.nonce.chars().any(char::is_control)
+        {
+            return Err(invalid_config());
+        }
+        Ok(())
+    }
+
+    pub fn from_path(path: &Path) -> Result<Self, RuntimeError> {
+        let mut options = OpenOptions::new();
+        options.read(true);
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::OpenOptionsExt;
+            // A FIFO must not stall startup before its regular-file check.
+            options.custom_flags(libc::O_NONBLOCK | libc::O_NOFOLLOW);
+        }
+        let file = options.open(path).map_err(|_| invalid_config())?;
+        let metadata = file.metadata().map_err(|_| invalid_config())?;
+        if !metadata.is_file() || metadata.len() > MAX_RESTORE_CHECKPOINT_CONFIG_BYTES as u64 {
+            return Err(invalid_config());
+        }
+        let mut bytes = Vec::new();
+        file.take(MAX_RESTORE_CHECKPOINT_CONFIG_BYTES as u64 + 1)
+            .read_to_end(&mut bytes)
+            .map_err(|_| invalid_config())?;
+        if bytes.len() > MAX_RESTORE_CHECKPOINT_CONFIG_BYTES {
+            return Err(invalid_config());
+        }
+        let config: Self = serde_json::from_slice(&bytes).map_err(|_| invalid_config())?;
+        config.validate()?;
+        Ok(config)
+    }
+}
+
+pub(crate) fn wait_for_release(config: &RestoreCheckpointConfig, ack: &RestoreCheckpointAck) {
+    let _ = exchange(config, ack);
+}
+
+fn remaining(deadline: Instant) -> std::io::Result<Duration> {
+    deadline
+        .checked_duration_since(Instant::now())
+        .filter(|remaining| !remaining.is_zero())
+        .ok_or_else(|| std::io::Error::from(std::io::ErrorKind::TimedOut))
+}
+
+fn exchange(config: &RestoreCheckpointConfig, ack: &RestoreCheckpointAck) -> std::io::Result<()> {
+    let deadline = Instant::now() + RESTORE_CHECKPOINT_TIMEOUT;
+    let mut bytes = serde_json::to_vec(ack)?;
+    bytes.push(b'\n');
+    if bytes.len() > MAX_RESTORE_CHECKPOINT_ACK_BYTES {
+        return Err(std::io::Error::from(std::io::ErrorKind::InvalidData));
+    }
+    let mut stream = TcpStream::connect_timeout(&config.endpoint, remaining(deadline)?)?;
+    // Each syscall shares one budget; partial writes cannot reset the deadline.
+    let mut pending = bytes.as_slice();
+    while !pending.is_empty() {
+        stream.set_write_timeout(Some(remaining(deadline)?))?;
+        let written = stream.write(pending)?;
+        if written == 0 {
+            return Err(std::io::Error::from(std::io::ErrorKind::WriteZero));
+        }
+        pending = &pending[written..];
+    }
+    stream.set_read_timeout(Some(remaining(deadline)?))?;
+    let mut release = [0];
+    stream.read_exact(&mut release)?;
+    if release != [b'R'] {
+        return Err(std::io::Error::from(std::io::ErrorKind::InvalidData));
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests;

crates/worker-runtime/src/restore_checkpoints/tests.rs

@@ -0,0 +1,139 @@
+use super::*;
+use std::io::{BufRead, BufReader};
+use std::net::TcpListener;
+use std::thread;
+
+fn config(endpoint: SocketAddr) -> RestoreCheckpointConfig {
+    RestoreCheckpointConfig {
+        endpoint,
+        worker_id: WorkerId::now_v7(),
+        stage: RestoreCheckpointStage::PreparedNotCommitted,
+        nonce: "checkpoint-test-handshake".into(),
+    }
+}
+
+fn ack(config: &RestoreCheckpointConfig) -> RestoreCheckpointAck {
+    RestoreCheckpointAck {
+        nonce: config.nonce.clone(),
+        stage: config.stage,
+        runtime_id: "checkpoint-runtime".into(),
+        worker_id: config.worker_id,
+        operation_id: WorkerLifecycleOperationId::new(),
+        commit: RestoreCheckpointCommit::NotCommitted,
+        receipt_absent: true,
+        pid: std::process::id(),
+    }
+}
+
+#[test]
+fn config_rejects_nonliteral_nonloopback_bad_stage_unknown_fields_and_invalid_worker() {
+    let valid = serde_json::to_value(config("127.0.0.1:1234".parse().unwrap())).unwrap();
+    for (field, value) in [
+        ("endpoint", "localhost:1234"),
+        ("endpoint", "192.0.2.1:1234"),
+        ("endpoint", "0.0.0.0:1234"),
+        ("endpoint", "127.0.0.1:0"),
+        ("endpoint", "[::ffff:127.0.0.1]:1234"),
+        ("worker_id", "not-a-worker-id"),
+        ("stage", "after_terminal"),
+        ("nonce", ""),
+        ("nonce", "secret\nbody"),
+        ("timeout", "unbounded"),
+    ] {
+        let mut invalid = valid.clone();
+        invalid[field] = value.into();
+        let parsed = serde_json::from_value::<RestoreCheckpointConfig>(invalid);
+        assert!(
+            parsed.is_err() || parsed.unwrap().validate().is_err(),
+            "{field}"
+        );
+    }
+    for endpoint in ["127.0.0.1:1234", "[::1]:1234"] {
+        config(endpoint.parse().unwrap()).validate().unwrap();
+    }
+}
+
+#[test]
+fn config_file_bounds_and_errors_do_not_disclose_contents() {
+    let dir = tempfile::tempdir().unwrap();
+    let path = dir.path().join("config.json");
+    assert!(RestoreCheckpointConfig::from_path(&path).is_err());
+    std::fs::write(&path, vec![b'x'; MAX_RESTORE_CHECKPOINT_CONFIG_BYTES + 1]).unwrap();
+    assert!(RestoreCheckpointConfig::from_path(&path).is_err());
+    std::fs::write(
+        &path,
+        br#"{"nonce":"private-config-body","stage":"secret"}"#,
+    )
+    .unwrap();
+    let error = RestoreCheckpointConfig::from_path(&path)
+        .unwrap_err()
+        .to_string();
+    assert!(!error.contains("private-config-body"));
+    assert!(!error.contains("secret"));
+    let mut valid = config("127.0.0.1:1234".parse().unwrap());
+    valid.nonce = "x".repeat(257);
+    assert!(valid.validate().is_err());
+    valid.nonce = "valid".into();
+    std::fs::write(&path, serde_json::to_vec(&valid).unwrap()).unwrap();
+    assert_eq!(RestoreCheckpointConfig::from_path(&path).unwrap(), valid);
+    assert!(RestoreCheckpointConfig::from_path(dir.path()).is_err());
+}
+
+#[test]
+fn explicit_release_eof_and_invalid_release_close_the_owned_stream() {
+    for release in [Some(b'R'), None, Some(b'X')] {
+        let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+        let config = config(listener.local_addr().unwrap());
+        let expected = ack(&config);
+        let sent = expected.clone();
+        let task = thread::spawn(move || exchange(&config, &sent));
+        let (mut stream, _) = listener.accept().unwrap();
+        stream
+            .set_read_timeout(Some(Duration::from_secs(2)))
+            .unwrap();
+        let mut line = String::new();
+        BufReader::new(&mut stream).read_line(&mut line).unwrap();
+        assert!(line.len() <= MAX_RESTORE_CHECKPOINT_ACK_BYTES);
+        assert_eq!(
+            serde_json::from_str::<RestoreCheckpointAck>(&line).unwrap(),
+            expected
+        );
+        if let Some(byte) = release {
+            stream.write_all(&[byte]).unwrap();
+        } else {
+            stream.shutdown(std::net::Shutdown::Write).unwrap();
+        }
+        assert_eq!(task.join().unwrap().is_ok(), release == Some(b'R'));
+        let mut byte = [0];
+        assert_eq!(stream.read(&mut byte).unwrap(), 0);
+    }
+}
+
+#[test]
+fn absent_listener_resumes_without_waiting_or_retrying() {
+    let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+    let config = config(listener.local_addr().unwrap());
+    drop(listener);
+    let start = Instant::now();
+    wait_for_release(&config, &ack(&config));
+    assert!(start.elapsed() < Duration::from_secs(2));
+}
+
+#[test]
+fn silent_peer_hits_fixed_deadline_and_stream_is_dropped() {
+    let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+    let config = config(listener.local_addr().unwrap());
+    let start = Instant::now();
+    let task = thread::spawn(move || exchange(&config, &ack(&config)));
+    let (mut stream, _) = listener.accept().unwrap();
+    stream
+        .set_read_timeout(Some(Duration::from_secs(12)))
+        .unwrap();
+    let mut reader = BufReader::new(&mut stream);
+    let mut line = String::new();
+    reader.read_line(&mut line).unwrap();
+    assert!(task.join().unwrap().is_err());
+    assert!(start.elapsed() >= RESTORE_CHECKPOINT_TIMEOUT - Duration::from_millis(100));
+    assert!(start.elapsed() < RESTORE_CHECKPOINT_TIMEOUT + Duration::from_secs(3));
+    assert_eq!(reader.read(&mut [0]).unwrap(), 0);
+}

crates/worker-runtime/src/retention/runtime_v5_tests.rs

@@ -3,7 +3,7 @@
 //! only the empty segment envelope is built with current APIs.
 use super::tests::{request, write_json};
 use super::*;
-use crate::fs_store::{FsRuntimeStore, FsRuntimeStoreOptions};
+use crate::fs_store::{FsRuntimeStore, FsRuntimeStoreOptions, SCHEMA_VERSION};
 use serde_json::{Value, json};
 use session_store::{LogEntry, SegmentId, SessionId, Store, WorkerSessionStore};
 use std::collections::BTreeMap;
@@ -226,7 +226,7 @@ fn completed_v4_archive(
 
 #[test]
 fn runtime_v11_and_current_route_v4_stopped_orphan_archive_through_staging() {
-    for runtime_schema in [11, 12] {
+    for runtime_schema in [11, 12, SCHEMA_VERSION] {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime");
         let (worker, session, segment) = runtime_fixture(&root, runtime_schema);
@@ -329,7 +329,7 @@ fn runtime_v11_and_current_route_v4_stopped_orphan_archive_through_staging() {
             (storage.checksum_sha256, storage.bytes, storage.file_count),
             (checksum, bytes, count)
         );
-        if runtime_schema == 12 {
+        if runtime_schema == SCHEMA_VERSION {
             for name in [
                 PathBuf::from("runtime.json"),
                 PathBuf::from(format!("workers/{worker}/worker.json")),
@@ -346,8 +346,8 @@ fn runtime_v11_and_current_route_v4_stopped_orphan_archive_through_staging() {
         );
         drop(FsRuntimeStore::open_or_create(root.clone(), "runtime-a").unwrap());
         assert_eq!(tree_bytes(&root), once, "retry rewrote migrated evidence");
-        assert!(!temp.path().join(".runtime.schema-v12-staging").exists());
-        assert!(!temp.path().join(".runtime.pre-schema-v12-backup").exists());
+        assert!(!temp.path().join(".runtime.schema-v13-staging").exists());
+        assert!(!temp.path().join(".runtime.pre-schema-v13-backup").exists());
         // session-storage, not the immutable receipt's old checksum, attests
         // current bytes. Tampering with those bytes must remain detectable.
         let path = archive_session
@@ -360,7 +360,7 @@ fn runtime_v11_and_current_route_v4_stopped_orphan_archive_through_staging() {
 
 #[test]
 fn corrupt_archive_does_not_block_startup_or_healthy_session_migration_and_repair_converges() {
-    for runtime_schema in [11, 12] {
+    for runtime_schema in [11, 12, SCHEMA_VERSION] {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime");
         let (worker, session, segment) = runtime_fixture(&root, runtime_schema);
@@ -440,8 +440,8 @@ fn unmigratable_archive_preserves_evidence_across_startup_and_reopen() {
     ] {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime");
-        let (live_worker, live_session, live_segment) = runtime_fixture(&root, 12);
-        let (worker, _, segment) = stopped_worker(&root, 12);
+        let (live_worker, live_session, live_segment) = runtime_fixture(&root, SCHEMA_VERSION);
+        let (worker, _, segment) = stopped_worker(&root, SCHEMA_VERSION);
         let provider = FsWorkerRetentionProvider::new(&root);
         let path = provider
             .worker_dir(worker)
@@ -542,8 +542,8 @@ fn unreadable_archive_metadata_does_not_block_archive_only_runtime_startup() {
     ] {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime");
-        let (worker, _, _) = runtime_fixture(&root, 12);
-        completed_v4_archive(&root, 12);
+        let (worker, _, _) = runtime_fixture(&root, SCHEMA_VERSION);
+        completed_v4_archive(&root, SCHEMA_VERSION);
         let provider = FsWorkerRetentionProvider::new(&root);
         fs::remove_dir_all(provider.worker_dir(worker)).unwrap();
         let archive = provider.archive_dir("archive-a").unwrap();
@@ -575,8 +575,8 @@ fn unreadable_archive_metadata_does_not_block_archive_only_runtime_startup() {
 fn archive_isolation_does_not_bypass_corrupt_erasure_authority() {
     let temp = tempfile::tempdir().unwrap();
     let root = temp.path().join("runtime");
-    runtime_fixture(&root, 12);
-    let (session, _, _, _) = completed_v4_archive(&root, 12);
+    runtime_fixture(&root, SCHEMA_VERSION);
+    let (session, _, _, _) = completed_v4_archive(&root, SCHEMA_VERSION);
     let ledger = root.join("retention/session-sources");
     fs::create_dir_all(&ledger).unwrap();
     fs::write(ledger.join(format!("{session}.json")), b"{broken\n").unwrap();
@@ -591,7 +591,7 @@ fn archive_isolation_does_not_bypass_corrupt_erasure_authority() {
 fn runtime_v5_probe_and_staging_skip_erased_originals() {
     let temp = tempfile::tempdir().unwrap();
     let root = temp.path().join("runtime");
-    let (worker, live_session, live_segment) = runtime_fixture(&root, 12);
+    let (worker, live_session, live_segment) = runtime_fixture(&root, SCHEMA_VERSION);
     let erased = root
         .join("orphaned-workers")
         .join(WorkerId::now_v7().to_string());
@@ -643,7 +643,7 @@ fn runtime_v5_rejects_broken_or_future_source_without_rewriting_evidence() {
     for corruption in ["broken-json", "future-schema"] {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime");
-        let (worker, session, segment) = runtime_fixture(&root, 12);
+        let (worker, session, segment) = runtime_fixture(&root, SCHEMA_VERSION);
         let session_root = root
             .join("workers")
             .join(worker.to_string())
@@ -674,19 +674,19 @@ fn runtime_v5_rejects_broken_or_future_source_without_rewriting_evidence() {
         assert_eq!(tree_bytes(&root), before);
         assert!(FsRuntimeStore::migrate(&options).is_err());
         assert_eq!(tree_bytes(&root), before);
-        assert!(!temp.path().join(".runtime.schema-v12-staging").exists());
-        assert!(!temp.path().join(".runtime.pre-schema-v12-backup").exists());
+        assert!(!temp.path().join(".runtime.schema-v13-staging").exists());
+        assert!(!temp.path().join(".runtime.pre-schema-v13-backup").exists());
     }
 }
 
 #[cfg(unix)]
 #[test]
 fn current_runtime_session_migration_still_rejects_live_worker_sockets() {
-    // Relative socket names keep this fixture independent of the checkout/TMPDIR
-    // length. The TempDir still owns the same tree; no process cwd is changed.
+    // The same TempDir owns the relative bind/connect reference. No cwd change
+    // or alias is needed, and the real live-socket migration refusal stays tested.
     let temp = tempfile::tempdir_in(".").unwrap();
     let root = Path::new(temp.path().file_name().unwrap()).join("runtime");
-    let (worker, _, _) = runtime_fixture(&root, 12);
+    let (worker, _, _) = runtime_fixture(&root, SCHEMA_VERSION);
     let before = tree_bytes(&root);
     let runs = root.join("workers").join(worker.to_string()).join("runs");
     fs::create_dir_all(runs.join("1")).unwrap();
@@ -697,7 +697,7 @@ fn current_runtime_session_migration_still_rejects_live_worker_sockets() {
     drop(listener);
     fs::remove_dir_all(runs).unwrap();
     assert_eq!(tree_bytes(&root), before);
-    assert!(!temp.path().join(".runtime.schema-v12-staging").exists());
+    assert!(!temp.path().join(".runtime.schema-v13-staging").exists());
 }
 
 #[test]
@@ -705,7 +705,7 @@ fn current_runtime_upgrades_v5_role_metadata_through_staging_without_reserializi
 {
     let temp = tempfile::tempdir().unwrap();
     let root = temp.path().join("runtime");
-    let (worker, session, segment) = runtime_fixture(&root, 12);
+    let (worker, session, segment) = runtime_fixture(&root, SCHEMA_VERSION);
     let aggregate = root.join("workers").join(worker.to_string());
     let path = aggregate
         .join("session/segments")

crates/worker-runtime/src/runtime/restore_checkpoint.rs

@@ -0,0 +1,112 @@
+use super::*;
+use crate::restore_checkpoints::{
+    RestoreCheckpointAck, RestoreCheckpointCommit, RestoreCheckpointConfig, RestoreCheckpointStage,
+    wait_for_release,
+};
+
+impl Runtime {
+    /// Install an explicitly selected, one-shot process-test gate after normal startup.
+    /// It is not persisted or consulted by admission, routing or recovery authority.
+    pub fn install_test_restore_checkpoint(
+        &self,
+        config: RestoreCheckpointConfig,
+    ) -> Result<(), RuntimeError> {
+        config.validate()?;
+        let mut state = self.lock()?;
+        if !matches!(state.persistence, RuntimePersistence::Fs(_)) {
+            return Err(RuntimeError::InvalidRequest(
+                "test restore checkpoint requires an fs Runtime store".into(),
+            ));
+        }
+        state.restore_checkpoint = Some(config);
+        Ok(())
+    }
+
+    // The lifecycle caller holds the per-Worker operation lock. No Runtime-state
+    // mutex is retained while waiting; the real backend candidate remains owned by
+    // the same operation throughout this synchronous call.
+    pub(super) fn test_restore_checkpoint(
+        &self,
+        worker_ref: &WorkerRef,
+        operation_id: WorkerLifecycleOperationId,
+        stage: RestoreCheckpointStage,
+    ) {
+        let config = {
+            let Ok(mut state) = self.lock() else {
+                return;
+            };
+            let slot = &mut state.restore_checkpoint;
+            if !slot.as_ref().is_some_and(|config| {
+                config.worker_id == worker_ref.worker_id && config.stage == stage
+            }) {
+                return;
+            }
+            // Consume even if canonical verification or the handshake fails.
+            slot.take().expect("matched checkpoint")
+        };
+        let Ok(ack) = self.restore_checkpoint_ack(&config, worker_ref, operation_id) else {
+            return;
+        };
+        wait_for_release(&config, &ack);
+    }
+
+    pub(super) fn restore_checkpoint_ack(
+        &self,
+        config: &RestoreCheckpointConfig,
+        worker_ref: &WorkerRef,
+        operation_id: WorkerLifecycleOperationId,
+    ) -> Result<RestoreCheckpointAck, RuntimeError> {
+        let invalid = || {
+            RuntimeError::InvalidRequest("test restore checkpoint authority did not match".into())
+        };
+        let state = self.lock()?;
+        let RuntimePersistence::Fs(store) = &state.persistence else {
+            return Err(invalid());
+        };
+        // The existing helper validates the exact bounded canonical aggregate and
+        // syncs that same file plus its owning directories. Never manufacture a
+        // journal marker or infer commit from an in-memory projection.
+        let record = store.read_durable_worker_record(worker_ref)?;
+        let PersistedWorkerExecutionState::ReconciliationRequired(
+            PersistedWorkerLifecycleOperation::Restore(pending),
+        ) = &record.execution_state
+        else {
+            return Err(invalid());
+        };
+        let expected = match config.stage {
+            RestoreCheckpointStage::PreparedNotCommitted => RestoreCommitFact::NotCommitted,
+            RestoreCheckpointStage::CommittedPreterminal
+            | RestoreCheckpointStage::ActivatedPreterminal => RestoreCommitFact::Committed,
+        };
+        if config.worker_id != worker_ref.worker_id
+            || pending.operation_id != operation_id
+            || pending.commit != expected
+        {
+            return Err(invalid());
+        }
+        let guard = &record.restore_guard;
+        let id = guard.active_request_id.as_ref().ok_or_else(invalid)?;
+        let owner = guard.owners.get(id).ok_or_else(invalid)?;
+        if owner.operation_id != operation_id
+            || owner.request.request_id != *id
+            || owner.receipt.is_some()
+        {
+            return Err(invalid());
+        }
+        let runtime_id = state.runtime_identity.clone().ok_or_else(invalid)?;
+        Ok(RestoreCheckpointAck {
+            nonce: config.nonce.clone(),
+            stage: config.stage,
+            runtime_id,
+            worker_id: record.worker_id,
+            operation_id,
+            commit: match expected {
+                RestoreCommitFact::NotCommitted => RestoreCheckpointCommit::NotCommitted,
+                RestoreCommitFact::Committed => RestoreCheckpointCommit::Committed,
+                RestoreCommitFact::LegacyUnknown => unreachable!(),
+            },
+            receipt_absent: true,
+            pid: std::process::id(),
+        })
+    }
+}

crates/worker-runtime/src/runtime/restore_checkpoint_tests.rs

@@ -0,0 +1,304 @@
+use super::*;
+use crate::restore_checkpoints::{
+    RestoreCheckpointAck, RestoreCheckpointCommit, RestoreCheckpointConfig, RestoreCheckpointStage,
+};
+use std::io::{BufRead, BufReader, Write};
+use std::net::TcpListener;
+use std::time::Duration;
+
+#[test]
+fn checkpoint_stages_ack_canonical_decision_and_eof_resumes_same_operation_once() {
+    // Component scope: durable Runtime with a substituted execution backend.
+    // Real prepared Controllers and process death are owned by the E2E suite.
+    for (stage, release) in [
+        (RestoreCheckpointStage::PreparedNotCommitted, "release"),
+        (RestoreCheckpointStage::CommittedPreterminal, "eof"),
+        (RestoreCheckpointStage::CommittedPreterminal, "timeout"),
+        (RestoreCheckpointStage::ActivatedPreterminal, "release"),
+    ] {
+        let dir = tempfile::tempdir().unwrap();
+        let backend = Arc::new(TestExecutionBackend::default());
+        let runtime = Runtime::with_fs_store_and_execution_backend(
+            FsRuntimeStoreOptions::new(dir.path().join("runtime"))
+                .with_runtime_id("checkpoint-runtime"),
+            backend.clone(),
+        )
+        .unwrap();
+        runtime.bind_runtime_identity("checkpoint-runtime").unwrap();
+        runtime.store_config_bundle(test_bundle()).unwrap();
+        let worker = runtime
+            .create_worker(task_request("checkpoint original operation"))
+            .unwrap();
+        runtime.stop_worker(&worker.worker_ref, None).unwrap();
+        let request = runtime.test_restore_request(&worker.worker_ref);
+        let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+        runtime
+            .install_test_restore_checkpoint(RestoreCheckpointConfig {
+                endpoint: listener.local_addr().unwrap(),
+                worker_id: worker.worker_id,
+                stage,
+                nonce: "runtime-stage-test".into(),
+            })
+            .unwrap();
+        let cloned = runtime.clone();
+        let worker_ref = worker.worker_ref.clone();
+        let sent_request = request.clone();
+        let (sent, received) = std::sync::mpsc::channel();
+        let task = std::thread::spawn(move || {
+            sent.send(cloned.restore_worker_operation(&worker_ref, sent_request))
+                .unwrap();
+        });
+        listener.set_nonblocking(true).unwrap();
+        let stream = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap()
+            .block_on(async {
+                let listener = tokio::net::TcpListener::from_std(listener).unwrap();
+                let (stream, _) = tokio::time::timeout(Duration::from_secs(5), listener.accept())
+                    .await
+                    .expect("checkpoint did not connect")
+                    .unwrap();
+                stream.into_std().unwrap()
+            });
+        stream.set_nonblocking(false).unwrap();
+        stream
+            .set_read_timeout(Some(Duration::from_secs(3)))
+            .unwrap();
+        let mut reader = BufReader::new(stream);
+        let mut line = String::new();
+        reader.read_line(&mut line).unwrap();
+        let ack: RestoreCheckpointAck = serde_json::from_str(&line).unwrap();
+        assert_eq!(ack.stage, stage);
+        assert_eq!(ack.nonce, "runtime-stage-test");
+        assert_eq!(ack.runtime_id, "checkpoint-runtime");
+        assert_eq!(ack.worker_id, worker.worker_id);
+        assert_eq!(ack.pid, std::process::id());
+        assert!(ack.receipt_absent);
+        assert_eq!(
+            ack.commit,
+            match stage {
+                RestoreCheckpointStage::PreparedNotCommitted =>
+                    RestoreCheckpointCommit::NotCommitted,
+                RestoreCheckpointStage::CommittedPreterminal
+                | RestoreCheckpointStage::ActivatedPreterminal => RestoreCheckpointCommit::Committed,
+            }
+        );
+        assert_eq!(
+            backend.restore_operation_ids.lock().unwrap().as_slice(),
+            &[ack.operation_id]
+        );
+        if stage == RestoreCheckpointStage::ActivatedPreterminal {
+            assert_eq!(backend.activation_operation_ids.lock().unwrap().as_slice(), &[ack.operation_id]);
+        } else {
+            assert!(backend.activation_operation_ids.lock().unwrap().is_empty());
+        }
+        assert!(received.try_recv().is_err());
+        let canonical = runtime_store(&runtime)
+            .read_durable_worker_record(&worker.worker_ref)
+            .unwrap();
+        let id = canonical.restore_guard.active_request_id.as_ref().unwrap();
+        assert_eq!(id, &request.request_id);
+        assert!(canonical.restore_guard.owners[id].receipt.is_none());
+        // Release, EOF and timeout all resume the original lifecycle owner.
+        let start = std::time::Instant::now();
+        if release == "release" {
+            reader.get_mut().write_all(b"R").unwrap();
+        } else if release == "eof" {
+            reader
+                .get_mut()
+                .shutdown(std::net::Shutdown::Write)
+                .unwrap();
+        }
+        let result = received
+            .recv_timeout(Duration::from_secs(12))
+            .unwrap()
+            .unwrap();
+        assert!(start.elapsed() < Duration::from_secs(11));
+        task.join().unwrap();
+        assert_eq!(result.state, WorkerRestoreState::Accepted);
+        assert_eq!(
+            backend.activation_operation_ids.lock().unwrap().as_slice(),
+            &[ack.operation_id]
+        );
+        assert_eq!(
+            runtime
+                .restore_worker_operation(&worker.worker_ref, request)
+                .unwrap(),
+            result
+        );
+        assert_eq!(*backend.restore_count.lock().unwrap(), 1);
+        assert!(runtime.lock().unwrap().restore_checkpoint.is_none());
+        runtime.stop_worker(&worker.worker_ref, None).unwrap();
+        let next = runtime.test_restore_request(&worker.worker_ref);
+        assert_eq!(
+            runtime
+                .restore_worker_operation(&worker.worker_ref, next)
+                .unwrap()
+                .state,
+            WorkerRestoreState::Accepted
+        );
+        assert_eq!(*backend.restore_count.lock().unwrap(), 2);
+    }
+}
+
+#[test]
+fn checkpoint_disabled_mismatched_or_unreachable_does_not_change_restore_result() {
+    for mode in ["disabled", "mismatch", "unreachable"] {
+        let dir = tempfile::tempdir().unwrap();
+        let backend = Arc::new(TestExecutionBackend::default());
+        let runtime = Runtime::with_fs_store_and_execution_backend(
+            FsRuntimeStoreOptions::new(dir.path().join("runtime"))
+                .with_runtime_id("checkpoint-runtime"),
+            backend.clone(),
+        )
+        .unwrap();
+        runtime.bind_runtime_identity("checkpoint-runtime").unwrap();
+        runtime.store_config_bundle(test_bundle()).unwrap();
+        let worker = runtime
+            .create_worker(task_request("checkpoint absent"))
+            .unwrap();
+        runtime.stop_worker(&worker.worker_ref, None).unwrap();
+        let request = runtime.test_restore_request(&worker.worker_ref);
+        if mode != "disabled" {
+            let listener = TcpListener::bind("127.0.0.1:0").unwrap();
+            let endpoint = listener.local_addr().unwrap();
+            drop(listener);
+            runtime
+                .install_test_restore_checkpoint(RestoreCheckpointConfig {
+                    endpoint,
+                    worker_id: if mode == "mismatch" {
+                        WorkerId::now_v7()
+                    } else {
+                        worker.worker_id
+                    },
+                    stage: RestoreCheckpointStage::PreparedNotCommitted,
+                    nonce: "absent-test".into(),
+                })
+                .unwrap();
+        }
+        let result = runtime
+            .restore_worker_operation(&worker.worker_ref, request.clone())
+            .unwrap();
+        assert_eq!(result.state, WorkerRestoreState::Accepted, "{mode}");
+        assert_eq!(
+            runtime
+                .restore_worker_operation(&worker.worker_ref, request)
+                .unwrap(),
+            result
+        );
+        assert_eq!(*backend.restore_count.lock().unwrap(), 1);
+        assert_eq!(
+            runtime.lock().unwrap().restore_checkpoint.is_some(),
+            mode == "mismatch"
+        );
+    }
+}
+
+#[test]
+fn checkpoint_verification_requires_canonical_worker_operation_commit_and_absent_receipt() {
+    let dir = tempfile::tempdir().unwrap();
+    let root = dir.path().join("runtime");
+    let backend = Arc::new(TestExecutionBackend::default());
+    let runtime = Runtime::with_fs_store_and_execution_backend(
+        FsRuntimeStoreOptions::new(&root).with_runtime_id("checkpoint-runtime"),
+        backend.clone(),
+    )
+    .unwrap();
+    runtime.bind_runtime_identity("checkpoint-runtime").unwrap();
+    runtime.store_config_bundle(test_bundle()).unwrap();
+    let worker = runtime
+        .create_worker(task_request("checkpoint validation"))
+        .unwrap();
+    runtime.stop_worker(&worker.worker_ref, None).unwrap();
+    let request = runtime.test_restore_request(&worker.worker_ref);
+    *backend.restore_result.lock().unwrap() = Some(WorkerExecutionSpawnResult::Errored(
+        WorkerExecutionResult::errored(WorkerExecutionOperation::Restore, "not yet connected"),
+    ));
+    assert_eq!(
+        runtime
+            .restore_worker_operation(&worker.worker_ref, request.clone())
+            .unwrap()
+            .state,
+        WorkerRestoreState::ReconciliationRequired
+    );
+    let operation = backend.restore_operation_ids.lock().unwrap()[0];
+    let mut config = RestoreCheckpointConfig {
+        endpoint: "127.0.0.1:1234".parse().unwrap(),
+        worker_id: worker.worker_id,
+        stage: RestoreCheckpointStage::PreparedNotCommitted,
+        nonce: "verification-test".into(),
+    };
+    let path = root
+        .join("workers")
+        .join(worker.worker_id.to_string())
+        .join("worker.json");
+    let before = std::fs::read(&path).unwrap();
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_ok()
+    );
+    assert!(
+        runtime
+            .restore_checkpoint_ack(
+                &config,
+                &worker.worker_ref,
+                WorkerLifecycleOperationId::new()
+            )
+            .is_err()
+    );
+    config.worker_id = WorkerId::now_v7();
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_err()
+    );
+    config.worker_id = worker.worker_id;
+    config.stage = RestoreCheckpointStage::CommittedPreterminal;
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_err()
+    );
+    config.stage = RestoreCheckpointStage::ActivatedPreterminal;
+    assert!(runtime.restore_checkpoint_ack(&config, &worker.worker_ref, operation).is_err());
+    config.stage = RestoreCheckpointStage::PreparedNotCommitted;
+    runtime_store(&runtime).fail_next_worker_reconcile_sync();
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_err()
+    );
+    assert_eq!(
+        std::fs::read(&path).unwrap(),
+        before,
+        "verification must not write authority"
+    );
+    // In-memory equality must not hide an unreadable canonical aggregate.
+    std::fs::write(&path, b"unreadable canonical aggregate").unwrap();
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_err()
+    );
+    assert_eq!(
+        std::fs::read(&path).unwrap(),
+        b"unreadable canonical aggregate"
+    );
+    std::fs::write(&path, before).unwrap();
+    backend.restore_result.lock().unwrap().take();
+    assert_eq!(
+        runtime
+            .restore_worker_operation(&worker.worker_ref, request)
+            .unwrap()
+            .state,
+        WorkerRestoreState::Accepted
+    );
+    assert!(
+        runtime
+            .restore_checkpoint_ack(&config, &worker.worker_ref, operation)
+            .is_err(),
+        "terminal operation cannot ack a preterminal checkpoint"
+    );
+}

crates/worker-runtime/src/runtime/restore_continuation.rs

@@ -0,0 +1,522 @@
+//! Runtime-owned continuation of an admitted restore, independent of its caller.
+//!
+//! Completion callbacks only signal a condition variable. They never acquire a
+//! Runtime/lifecycle lock and are never joined by this driver while holding one.
+//! Each step revalidates the canonical owner under the existing Worker lock;
+//! waiting/backoff and driver teardown happen outside it. A repeated obstruction
+//! parks this owner with its reason, rather than spinning or inventing rollback.
+
+use super::*;
+use crate::execution::RestoreRecoveryDisposition;
+use std::collections::HashMap;
+use std::sync::{Condvar, Weak};
+use std::thread::JoinHandle;
+use std::time::Duration;
+
+type Key = (WorkerRef, WorkerLifecycleOperationId);
+
+#[derive(Debug, Default)]
+pub(super) struct RestoreContinuations {
+    owners: Mutex<HashMap<Key, Arc<Continuation>>>,
+    #[cfg(test)]
+    paused: std::sync::atomic::AtomicBool,
+    #[cfg(test)]
+    pub(super) terminal_receipt_gate:
+        Mutex<Option<crate::execution::RestoreTestGate<WorkerLifecycleOperationId>>>,
+}
+
+#[derive(Debug, Default)]
+struct Schedule {
+    signaled: bool,
+    cancelled: bool,
+    finished: bool,
+    blocked_reason: Option<String>,
+    parked: bool,
+}
+
+#[derive(Debug, Default)]
+struct Continuation {
+    schedule: Mutex<Schedule>,
+    wake: Condvar,
+    task: Mutex<Option<JoinHandle<()>>>,
+}
+
+impl RestoreContinuations {
+    pub(super) fn signal(&self, worker_ref: &WorkerRef, operation: WorkerLifecycleOperationId) {
+        if let Some(owner) = self
+            .owners
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .get(&(worker_ref.clone(), operation))
+        {
+            let mut schedule = owner
+                .schedule
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            schedule.signaled = true;
+            owner.wake.notify_all();
+        }
+    }
+}
+
+impl Drop for RestoreContinuations {
+    fn drop(&mut self) {
+        let owners = self
+            .owners
+            .get_mut()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        for owner in owners.values() {
+            let mut schedule = owner
+                .schedule
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            schedule.cancelled = true;
+            owner.wake.notify_all();
+        }
+        for owner in owners.values() {
+            if let Some(task) = owner
+                .task
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner)
+                .take()
+            {
+                if task.thread().id() != std::thread::current().id() {
+                    let _ = task.join();
+                }
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+pub(crate) struct ContinuationPause(Arc<RestoreContinuations>);
+
+#[cfg(test)]
+impl Drop for ContinuationPause {
+    fn drop(&mut self) {
+        self.0
+            .paused
+            .store(false, std::sync::atomic::Ordering::Release);
+        for owner in self.0.owners.lock().unwrap().values() {
+            owner.schedule.lock().unwrap().signaled = true;
+            owner.wake.notify_all();
+        }
+    }
+}
+
+enum Step {
+    Finished,
+    Waiting,
+    Contended,
+    Retry(String),
+}
+
+impl Runtime {
+    pub(super) fn restore_repair_authorizer(
+        &self,
+        worker_ref: WorkerRef,
+    ) -> Arc<dyn Fn(WorkerLifecycleOperationId) -> Result<(), RuntimeError> + Send + Sync> {
+        let inner = Arc::downgrade(&self.inner);
+        Arc::new(move |operation| {
+            let inner = inner.upgrade().ok_or(RuntimeError::RuntimeStopped)?;
+            // The factory caller owns the lifecycle lock. Never reacquire it
+            // here: it may be awaiting this exact factory task.
+            let mut state = inner.lock().map_err(|_| RuntimeError::StatePoisoned)?;
+            state.ensure_running()?;
+            state.refresh_uncertain_worker_write(&worker_ref)?;
+            let worker = state.worker(&worker_ref)?;
+            state.ensure_subject_not_reserved(&worker_ref, false)?;
+            if worker.pending_stop.is_some()
+                || worker.pending_restore.is_none_or(|pending| {
+                    pending.operation_id != operation
+                        || pending.commit == RestoreCommitFact::LegacyUnknown
+                })
+            {
+                return Err(RuntimeError::RestoreObservationConflict {
+                    worker_id: worker_ref.worker_id,
+                });
+            }
+            if let Some(owner) = worker
+                .restore_guard
+                .active_request_id
+                .as_ref()
+                .and_then(|id| worker.restore_guard.owners.get(id))
+            {
+                owner.validate_workdir_preparation_authority(worker)?;
+            }
+            if worker
+                .pending_restore
+                .is_some_and(|pending| pending.source_repair_possible)
+            {
+                return Ok(());
+            }
+            let mut candidate = worker.clone();
+            candidate
+                .pending_restore
+                .as_mut()
+                .expect("validated restore owner")
+                .source_repair_possible = true;
+            // Grant before touching another canonical store. Unknown durability
+            // denies repair until authoritative reconciliation; no old-memory write.
+            state.persist_restore_worker_record(&candidate)?;
+            state.workers.insert(worker_ref.worker_id, candidate);
+            Ok(())
+        })
+    }
+
+    #[cfg(test)]
+    pub(crate) fn test_wait_restore_continuation_parked(
+        &self,
+        worker_ref: &WorkerRef,
+        operation: WorkerLifecycleOperationId,
+        timeout: Duration,
+    ) -> String {
+        let owner = self
+            .restore_continuations
+            .owners
+            .lock()
+            .unwrap()
+            .get(&(worker_ref.clone(), operation))
+            .cloned()
+            .unwrap();
+        let (schedule, waited) = owner
+            .wake
+            .wait_timeout_while(owner.schedule.lock().unwrap(), timeout, |s| {
+                !s.parked && !s.finished
+            })
+            .unwrap();
+        assert!(
+            !waited.timed_out() && schedule.parked,
+            "continuation did not park: {:?}",
+            schedule.blocked_reason
+        );
+        schedule.blocked_reason.clone().expect("parked reason")
+    }
+
+    /// Gate autonomous steps when a component test observes a single fault
+    /// boundary. Dedicated continuation tests run without this gate.
+    #[cfg(test)]
+    pub(crate) fn test_pause_restore_continuation(&self) -> ContinuationPause {
+        assert!(
+            !self
+                .restore_continuations
+                .paused
+                .swap(true, std::sync::atomic::Ordering::AcqRel)
+        );
+        ContinuationPause(self.restore_continuations.clone())
+    }
+
+    #[cfg(test)]
+    pub(crate) fn test_wait_restore_continuation(
+        &self,
+        worker_ref: &WorkerRef,
+        operation: WorkerLifecycleOperationId,
+        timeout: Duration,
+    ) {
+        let owner = self
+            .restore_continuations
+            .owners
+            .lock()
+            .unwrap()
+            .get(&(worker_ref.clone(), operation))
+            .cloned()
+            .expect("continuation owner");
+        let schedule = owner.schedule.lock().unwrap();
+        let (schedule, waited) = owner
+            .wake
+            .wait_timeout_while(schedule, timeout, |s| !s.finished && !s.cancelled)
+            .unwrap();
+        assert!(
+            !waited.timed_out() && schedule.finished,
+            "restore continuation did not settle: {:?}",
+            schedule.blocked_reason
+        );
+    }
+
+    pub(super) fn ensure_restore_continuation(
+        &self,
+        worker_ref: &WorkerRef,
+    ) -> Result<(), RuntimeError> {
+        let operation = {
+            let state = self.lock()?;
+            let worker = state.worker(worker_ref)?;
+            worker
+                .pending_restore
+                .map(|pending| pending.operation_id)
+                .or_else(|| {
+                    worker
+                        .restore_guard
+                        .active_request_id
+                        .as_ref()
+                        .and_then(|id| worker.restore_guard.owners.get(id))
+                        .filter(|owner| owner.receipt.is_none())
+                        .map(|owner| owner.operation_id)
+                })
+        };
+        if let Some(operation) = operation {
+            self.start_restore_continuation(worker_ref.clone(), operation)?;
+        }
+        Ok(())
+    }
+
+    pub(super) fn start_restore_continuation(
+        &self,
+        worker_ref: WorkerRef,
+        operation: WorkerLifecycleOperationId,
+    ) -> Result<(), RuntimeError> {
+        {
+            let state = self.lock()?;
+            let Some(backend) = &state.execution_backend else {
+                return Ok(());
+            };
+            if !backend.supports_restore_continuation()
+                || backend
+                    .restore_recovery_disposition(operation, &worker_ref)
+                    .is_none()
+            {
+                return Ok(());
+            }
+        }
+        let key = (worker_ref.clone(), operation);
+        let mut owners = self
+            .restore_continuations
+            .owners
+            .lock()
+            .map_err(|_| RuntimeError::StatePoisoned)?;
+        if owners.contains_key(&key) {
+            return Ok(());
+        }
+        let owner = Arc::new(Continuation::default());
+        let retained = owner.clone();
+        let state = Arc::downgrade(&self.inner);
+        let operations = Arc::downgrade(&self.worker_operations);
+        let continuations = Arc::downgrade(&self.restore_continuations);
+        let task = std::thread::Builder::new()
+            .name("restore-continuation".into())
+            .spawn(move || {
+                run(
+                    state,
+                    operations,
+                    continuations,
+                    retained,
+                    worker_ref,
+                    operation,
+                );
+            })
+            .map_err(|error| {
+                RuntimeError::InvalidRequest(format!(
+                    "Restore continuation could not start: {error}"
+                ))
+            })?;
+        *owner.task.lock().map_err(|_| RuntimeError::StatePoisoned)? = Some(task);
+        owners.insert(key, owner);
+        Ok(())
+    }
+
+    fn continue_restore_operation(
+        &self,
+        worker_ref: &WorkerRef,
+        operation: WorkerLifecycleOperationId,
+    ) -> Result<Step, RuntimeError> {
+        #[cfg(test)]
+        if self
+            .restore_continuations
+            .paused
+            .load(std::sync::atomic::Ordering::Acquire)
+        {
+            return Ok(Step::Contended);
+        }
+        let lease = self.worker_operation_lock(worker_ref.worker_id)?;
+        // Another caller may be waiting for a backend task. Do not queue a join
+        // of that caller/driver while retaining the lock it needs to complete.
+        let _guard = match lease.try_lock() {
+            Ok(guard) => guard,
+            Err(std::sync::TryLockError::WouldBlock) => return Ok(Step::Contended),
+            Err(_) => return Err(RuntimeError::StatePoisoned),
+        };
+        let stop_is_owned = {
+            let mut state = self.lock()?;
+            state.ensure_running()?;
+            state.refresh_uncertain_worker_write(worker_ref)?;
+            let worker = state.worker(worker_ref)?;
+            worker.pending_stop.is_some()
+                && worker
+                    .pending_restore
+                    .is_some_and(|pending| pending.operation_id == operation)
+        };
+        if stop_is_owned {
+            // Stop supersedes publication, not ownership of late resources.
+            // Retain this same continuation until its cleanup is confirmed.
+            self.reconcile_pending_stop_under_lock(worker_ref)?;
+            return Ok(Step::Finished);
+        }
+        let (backend, pending) = {
+            let mut state = self.lock()?;
+            state.ensure_running()?;
+            state.refresh_uncertain_worker_write(worker_ref)?;
+            let worker = state.worker(worker_ref)?;
+            if worker.pending_stop.is_some() {
+                return Ok(Step::Finished);
+            }
+            let owner = worker
+                .restore_guard
+                .active_request_id
+                .as_ref()
+                .and_then(|id| worker.restore_guard.owners.get(id));
+            if owner.is_some_and(|owner| owner.operation_id != operation || owner.receipt.is_some())
+            {
+                return Ok(Step::Finished);
+            }
+            let Some(pending) = worker.pending_restore else {
+                return Ok(Step::Finished);
+            };
+            if pending.operation_id != operation {
+                return Ok(Step::Finished);
+            }
+            if pending.commit == RestoreCommitFact::LegacyUnknown {
+                return Ok(Step::Retry(
+                    "Saved Restore has no authoritative commit evidence".into(),
+                ));
+            }
+            state.ensure_subject_not_reserved(worker_ref, false)?;
+            if let Some(owner) = owner {
+                owner.validate_workdir_preparation_authority(worker)?;
+            }
+            let Some(backend) = state.execution_backend.clone() else {
+                return Ok(Step::Finished);
+            };
+            (backend, pending)
+        };
+        match backend.restore_recovery_disposition(operation, worker_ref) {
+            None => return Ok(Step::Finished),
+            Some(RestoreRecoveryDisposition::InFlight) => return Ok(Step::Waiting),
+            Some(RestoreRecoveryDisposition::Discard)
+                if pending.commit == RestoreCommitFact::NotCommitted =>
+            {
+                match backend.discard_uncommitted_restore(operation, worker_ref) {
+                    WorkerExecutionSpawnResult::RolledBack(result) => {
+                        let receipt = RuntimeWorkerRestoreResult::failed(
+                            WorkerRestoreState::RolledBack,
+                            "worker_restore_rolled_back",
+                            result.message_or_default(),
+                        );
+                        #[cfg(feature = "fs-store")]
+                        self.lock()?
+                            .record_restore_failure(worker_ref, &result, &receipt)?;
+                        #[cfg(not(feature = "fs-store"))]
+                        self.settle_restore_rollback(worker_ref, &receipt)?;
+                        return Ok(Step::Finished);
+                    }
+                    result => {
+                        return Ok(Step::Retry(format!(
+                            "Restore cleanup remains unresolved: {result:?}"
+                        )));
+                    }
+                }
+            }
+            Some(_) => {}
+        }
+        let result = self.restore_admitted_worker_under_lock(worker_ref, pending.mode)?;
+        if backend.restore_recovery_disposition(operation, worker_ref)
+            == Some(RestoreRecoveryDisposition::InFlight)
+        {
+            return Ok(Step::Waiting);
+        }
+        if result.state == WorkerRestoreState::ReconciliationRequired {
+            Ok(Step::Retry(result.message.unwrap_or_else(|| {
+                "Restore reconciliation remains pending".into()
+            })))
+        } else {
+            Ok(Step::Finished)
+        }
+    }
+}
+
+fn run(
+    inner: Weak<Mutex<RuntimeState>>,
+    operations: Weak<Mutex<BTreeMap<WorkerId, Arc<Mutex<()>>>>>,
+    continuations: Weak<RestoreContinuations>,
+    owner: Arc<Continuation>,
+    worker_ref: WorkerRef,
+    operation: WorkerLifecycleOperationId,
+) {
+    let mut repeated = 0u32;
+    let mut last_reason = None;
+    loop {
+        let (Some(inner), Some(worker_operations), Some(restore_continuations)) = (
+            inner.upgrade(),
+            operations.upgrade(),
+            continuations.upgrade(),
+        ) else {
+            break;
+        };
+        owner
+            .schedule
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .parked = false;
+        let runtime = Runtime {
+            inner,
+            worker_operations,
+            restore_continuations,
+        };
+        let step = runtime.continue_restore_operation(&worker_ref, operation);
+        // No lifecycle lock or strong Runtime/registry reference survives waiting.
+        drop(runtime);
+        let mut schedule = owner
+            .schedule
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if schedule.cancelled {
+            break;
+        }
+        let contended = matches!(&step, Ok(Step::Contended));
+        let reason = match step {
+            Ok(Step::Finished) => {
+                schedule.finished = true;
+                owner.wake.notify_all();
+                break;
+            }
+            Ok(Step::Waiting) | Ok(Step::Contended) => None,
+            Ok(Step::Retry(reason)) => Some(reason),
+            Err(error) => Some(sanitize_worker_create_failure_message(&error.to_string())),
+        };
+        if let Some(reason) = &reason {
+            repeated = repeated.saturating_add(1);
+            last_reason = Some(reason.clone());
+            schedule.blocked_reason = Some(reason.clone());
+        }
+        // Completed failed attempts can have signaled while this step joined
+        // them. Do not let our own unchanged failure reset the retry budget.
+        if reason.is_some() {
+            schedule.signaled = false;
+        }
+        if (!contended && reason.is_none()) || repeated >= 4 {
+            if repeated >= 4 {
+                schedule.parked = true;
+                owner.wake.notify_all();
+                tracing::warn!(worker_id = %worker_ref.worker_id, operation_id = %operation, reason = ?last_reason,
+                    "Restore continuation parked after bounded attempts; operation remains unresolved");
+            }
+            schedule = owner
+                .wake
+                .wait_while(schedule, |s| !s.signaled && !s.cancelled)
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+        } else {
+            let delay = Duration::from_millis(100 * (1u64 << repeated.min(4)));
+            schedule = owner
+                .wake
+                .wait_timeout_while(schedule, delay, |s| !s.signaled && !s.cancelled)
+                .unwrap_or_else(std::sync::PoisonError::into_inner)
+                .0;
+        }
+        if schedule.cancelled {
+            break;
+        }
+        if schedule.signaled {
+            schedule.signaled = false;
+            // A completion wakes this same owner, not a new request. Its next
+            // observed result, not the signal itself, determines progress.
+        }
+    }
+}

crates/worker-runtime/src/test_support.rs

@@ -0,0 +1,872 @@
+//! Offline support for the approved, attachment-free stopped Worker fixture.
+//!
+//! Enabled only by `test-support`. This is not an import API: extraction accepts a
+//! current-schema, settled, single-Segment aggregate; installation accepts only a
+//! fresh Runtime directory containing its newly initialized `auth.toml`. No
+//! Runtime or execution backend is constructed. Original Session bytes are never
+//! rewritten. The caller must separately review/attest the artificial conversation;
+//! structural checks and secret-marker checks cannot prove arbitrary prose safe.
+
+use super::*;
+use crate::catalog::ProfileSourceArchiveSource;
+use crate::workspace_issuer::{
+    WorkspaceIssuerTrustRecord, WorkspaceIssuerTrustState, validate_workspace_issuer_trust_records,
+};
+use session_store::{
+    SegmentId, SessionId, WorkerActiveSegmentRef, WorkerMetadata, WorkerSessionStore,
+};
+
+#[cfg(test)]
+#[path = "test_support_tests.rs"]
+mod tests;
+
+const TEMPLATE_VERSION: u32 = 1;
+const MAX_SESSION_BYTES: usize = 8 * 1024 * 1024;
+
+/// Portable fixture, intentionally without Debug (Session material is private).
+/// Fields are private and every deserialized template is validated on install.
+/// Serialize this value to retain the fixture; do not serialize its source aggregate.
+#[derive(Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct StoppedWorkerTemplate {
+    template_version: u32,
+    runtime_schema_version: u32,
+    runtime_id: String,
+    workspace_id: String,
+    worker_id: WorkerId,
+    created_at_ms: Option<u64>,
+    session_id: SessionId,
+    segment_id: SegmentId,
+    session_manifest: Vec<u8>,
+    segment_jsonl: Vec<u8>,
+}
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+struct SessionManifest {
+    schema_version: u32,
+    session_id: SessionId,
+}
+
+// Read-only projection: never retain source key material in a template.
+#[derive(Deserialize)]
+struct AuthIdentity {
+    identity_id: String,
+}
+
+#[derive(Deserialize)]
+struct SourceAuth {
+    identity: AuthIdentity,
+}
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+struct InitializedAuth {
+    identity: crate::auth::RuntimeIdentityMaterial,
+    workspace_issuers: Vec<WorkspaceIssuerTrustRecord>,
+    #[serde(default)]
+    archived_legacy_issuers: Vec<String>,
+}
+
+#[derive(Deserialize)]
+struct GuardState {
+    active_request_id: Option<String>,
+    owners: BTreeMap<String, GuardOwner>,
+}
+
+#[derive(Deserialize)]
+struct GuardOwner {
+    receipt: Option<serde::de::IgnoredAny>,
+}
+
+fn reject(message: &'static str) -> RuntimeError {
+    RuntimeError::InvalidRequest(format!("stopped Worker fixture: {message}"))
+}
+
+fn io<T>(result: std::io::Result<T>) -> Result<T, RuntimeError> {
+    result.map_err(|_| reject("filesystem operation failed"))
+}
+
+fn decode<T: serde::de::DeserializeOwned>(bytes: &[u8]) -> Result<T, RuntimeError> {
+    serde_json::from_slice(bytes).map_err(|_| reject("invalid typed document"))
+}
+
+fn no_symlink_components(path: &Path) -> Result<(), RuntimeError> {
+    let mut current = PathBuf::new();
+    for part in path.components() {
+        if matches!(part, std::path::Component::ParentDir) {
+            return Err(reject("parent traversal is unsupported"));
+        }
+        current.push(part);
+        let metadata = io(fs::symlink_metadata(&current))?;
+        if metadata.file_type().is_symlink() {
+            return Err(reject("symlinks are unsupported"));
+        }
+    }
+    Ok(())
+}
+
+fn check_tree(path: &Path, budget: &mut usize) -> Result<(), RuntimeError> {
+    if *budget == 0 {
+        return Err(reject("file count exceeds fixture limit"));
+    }
+    *budget -= 1;
+    let metadata = io(fs::symlink_metadata(path))?;
+    if metadata.is_dir() {
+        for entry in io(fs::read_dir(path))? {
+            check_tree(&io(entry)?.path(), budget)?;
+        }
+    } else if !metadata.is_file() {
+        return Err(reject("symlinks and special files are unsupported"));
+    }
+    Ok(())
+}
+
+fn regular_file(path: &Path) -> Result<(), RuntimeError> {
+    no_symlink_components(path)?;
+    if !io(fs::symlink_metadata(path))?.is_file() {
+        return Err(reject("expected a regular file"));
+    }
+    Ok(())
+}
+
+fn bytes(path: &Path) -> Result<Vec<u8>, RuntimeError> {
+    regular_file(path)?;
+    let mut options = OpenOptions::new();
+    options.read(true);
+    #[cfg(target_os = "linux")]
+    {
+        use std::os::unix::fs::OpenOptionsExt;
+        options.custom_flags(libc::O_NOATIME | libc::O_NOFOLLOW);
+    }
+    let file = io(options.open(path))?;
+    let mut result = Vec::new();
+    io(file
+        .take(MAX_SESSION_BYTES as u64 + 1)
+        .read_to_end(&mut result))?;
+    if result.len() > MAX_SESSION_BYTES {
+        return Err(reject("file exceeds fixture limit"));
+    }
+    Ok(result)
+}
+
+fn text(path: &Path) -> Result<String, RuntimeError> {
+    String::from_utf8(bytes(path)?).map_err(|_| reject("invalid UTF-8 document"))
+}
+
+fn portable_id(id: &str) -> bool {
+    !id.is_empty()
+        && id.len() <= 256
+        && id
+            .bytes()
+            .all(|c| c.is_ascii_alphanumeric() || b"-_".contains(&c))
+}
+
+impl StoppedWorkerTemplate {
+    pub fn runtime_id(&self) -> &str {
+        &self.runtime_id
+    }
+    pub fn workspace_id(&self) -> &str {
+        &self.workspace_id
+    }
+    pub fn worker_id(&self) -> WorkerId {
+        self.worker_id
+    }
+    pub fn worker_name(&self) -> String {
+        runtime_worker_name(self.worker_id)
+    }
+    pub fn session_id(&self) -> SessionId {
+        self.session_id
+    }
+    pub fn segment_id(&self) -> SegmentId {
+        self.segment_id
+    }
+
+    /// Extract without migration, rewriting metadata, or copying credentials,
+    /// resolved manifests, execution-operation receipts, or run directories.
+    /// The source Runtime must be offline; its existing owner lock is held for
+    /// the entire read. The path must be `<runtime>/workers/<WorkerId>`.
+    pub fn extract(worker_dir: &Path) -> Result<Self, RuntimeError> {
+        no_symlink_components(worker_dir)?;
+        let workers = worker_dir
+            .parent()
+            .ok_or_else(|| reject("invalid aggregate path"))?;
+        if workers.file_name().and_then(|s| s.to_str()) != Some(WORKERS_DIR) {
+            return Err(reject("invalid aggregate path"));
+        }
+        let root = workers
+            .parent()
+            .ok_or_else(|| reject("invalid Runtime path"))?;
+        // Require the existing lock, rather than creating a file in the source.
+        let lock_path = root.with_file_name(format!(
+            ".{}.runtime-owner.lock",
+            root.file_name()
+                .and_then(|s| s.to_str())
+                .ok_or_else(|| reject("invalid Runtime path"))?
+        ));
+        regular_file(&lock_path)?;
+        let lock = io(OpenOptions::new().read(true).open(lock_path))?;
+        if !io(FileExt::try_lock_shared(&lock))? {
+            return Err(reject("source Runtime is not offline"));
+        }
+        let _lock = RuntimeStoreOwnerLock { file: lock };
+        check_tree(worker_dir, &mut 256)?;
+        let snapshot: RuntimeSnapshot = decode(&bytes(&root.join(RUNTIME_FILE))?)?;
+        snapshot
+            .validate(&root.join(RUNTIME_FILE))
+            .map_err(|_| reject("unsupported Runtime schema"))?;
+        if !snapshot.subject_create_recovery_fences.is_empty()
+            || !snapshot.subject_releases.is_empty()
+        {
+            return Err(reject("source has lifecycle operations"));
+        }
+        let record: WorkerAggregateRecord = decode(&bytes(&worker_dir.join(WORKER_FILE))?)?;
+        let record = record
+            .validate(&worker_dir.join(WORKER_FILE))
+            .map_err(|_| reject("invalid Worker schema or identity"))?;
+        let guard: GuardState = decode(
+            &serde_json::to_vec(&record.restore_guard)
+                .map_err(|_| reject("invalid restore guard"))?,
+        )?;
+        if record.status != WorkerStatus::Stopped
+            || guard.active_request_id.is_some()
+            || guard.owners.values().any(|owner| owner.receipt.is_none())
+        {
+            return Err(reject("Worker is not settled and stopped"));
+        }
+        let PersistedWorkerExecutionState::Available(execution) = &record.execution_state else {
+            return Err(reject("pending or unavailable execution metadata"));
+        };
+        // A stopped aggregate may retain the empty historical binding marker.
+        // It is deliberately not copied to the installed Worker.
+        if execution.restore_intent != WorkerRestoreIntent::Explicit {
+            return Err(reject("Worker requires explicit restore"));
+        }
+        if !record.workdir_attachments.is_empty() || !record.logical_workdir_attachments.is_empty()
+        {
+            return Err(reject("attachments are unsupported"));
+        }
+        check_request(&execution.request, false)?;
+        let workspace_id = record
+            .workspace_id
+            .ok_or_else(|| reject("missing Workspace identity"))?;
+        let metadata: WorkerMetadata = decode(&bytes(&worker_dir.join(WORKER_METADATA_FILE))?)?;
+        crate::worker_backend::validate_fixture_restore_metadata(&metadata, &execution.request)
+            .map_err(|_| reject("invalid source restore Manifest or settings"))?;
+        if metadata.worker_name != runtime_worker_name(record.worker_id)
+            || metadata.workspace_id.as_deref() != Some(&workspace_id)
+            || !metadata.spawned_children.is_empty()
+            || !metadata.reclaimed_children.is_empty()
+            || !metadata.peers.is_empty()
+            || metadata.subjektiv_session_attribution.is_some()
+        {
+            return Err(reject("unsupported or inconsistent Worker metadata"));
+        }
+        let active = metadata
+            .active
+            .ok_or_else(|| reject("missing active Session"))?;
+        let segment_id = active
+            .segment_id
+            .ok_or_else(|| reject("unmaterialized Session"))?;
+        let source_auth: SourceAuth = toml::from_str(&text(&root.join("auth.toml"))?)
+            .map_err(|_| reject("missing Runtime identity"))?;
+        let session = worker_dir.join("session");
+        let fixture = Self {
+            template_version: TEMPLATE_VERSION,
+            runtime_schema_version: SCHEMA_VERSION,
+            runtime_id: source_auth.identity.identity_id,
+            workspace_id,
+            worker_id: record.worker_id,
+            created_at_ms: record.created_at_ms,
+            session_id: active.session_id,
+            segment_id,
+            session_manifest: bytes(&session.join("session.json"))?,
+            segment_jsonl: bytes(&session.join("segments").join(format!("{segment_id}.jsonl")))?,
+        };
+        fixture.validate_shape()?;
+        fixture.validate_session(&session)?;
+        Ok(fixture)
+    }
+
+    /// Install after normal identity/trust initialization, before Runtime start.
+    /// `request` and `manifest` supply NEW backend/config/profile references and
+    /// freshly resolved policy, not a launch. The saved source Manifest is never
+    /// reused. `manifest.worker.name` must equal `worker_name()`.
+    /// Only `auth.toml` may already exist. The owner lock fences live Runtimes;
+    /// staged files are synced and published without overwriting any Worker.
+    /// An interrupted publication fails closed as an incomplete Runtime layout.
+    pub fn install(
+        &self,
+        options: &FsRuntimeStoreOptions,
+        request: CreateWorkerRequest,
+        manifest: &manifest::WorkerManifest,
+    ) -> Result<(), RuntimeError> {
+        self.validate_shape()?;
+        check_request(&request, true)?;
+        if manifest.worker.name != self.worker_name() {
+            return Err(reject("resolved Manifest Worker name mismatch"));
+        }
+        let mut metadata = WorkerMetadata::new(
+            self.worker_name(),
+            Some(WorkerActiveSegmentRef {
+                session_id: self.session_id,
+                segment_id: Some(self.segment_id),
+            }),
+        )
+        .with_workspace_id(&self.workspace_id);
+        metadata.resolved_manifest_snapshot = Some(
+            manifest::write_persisted_worker_manifest_snapshot(manifest)
+                .map_err(|_| reject("invalid resolved Manifest"))?,
+        );
+        crate::worker_backend::validate_fixture_restore_metadata(&metadata, &request)
+            .map_err(|_| reject("resolved Manifest does not match fresh request settings"))?;
+        if options.runtime_id != self.runtime_id
+            || request.worker_id != self.worker_id
+            || request
+                .workspace_api
+                .as_ref()
+                .map(|api| api.workspace_id.as_str())
+                != Some(&self.workspace_id)
+        {
+            return Err(reject("installation identity mismatch"));
+        }
+        no_symlink_components(&options.root)?;
+        check_tree(&options.root, &mut 256)?;
+        // Validate the lock location before the ordinary owner-lock opener.
+        let lock_path = options.root.with_file_name(format!(
+            ".{}.runtime-owner.lock",
+            options
+                .root
+                .file_name()
+                .and_then(|s| s.to_str())
+                .ok_or_else(|| reject("invalid Runtime path"))?
+        ));
+        if fs::symlink_metadata(&lock_path).is_ok() {
+            regular_file(&lock_path)?;
+        }
+        let lock = acquire_runtime_store_owner_lock(&options.root)?;
+        validate_fresh_runtime_store_layout(&options.root)
+            .map_err(|_| reject("destination is not fresh; refusing overwrite"))?;
+        let auth: InitializedAuth = toml::from_str(&text(&options.root.join("auth.toml"))?)
+            .map_err(|_| reject("initialize Runtime identity and trust before install"))?;
+        if auth.identity.identity_id != self.runtime_id
+            || !auth.archived_legacy_issuers.is_empty()
+            || auth.identity.signing_key().is_err()
+        {
+            return Err(reject("invalid destination Runtime identity"));
+        }
+        validate_workspace_issuer_trust_records(&auth.workspace_issuers)
+            .map_err(|_| reject("invalid destination Workspace trust"))?;
+        let api = request
+            .workspace_api
+            .as_ref()
+            .expect("validated Workspace binding");
+        if !auth.workspace_issuers.iter().any(|trust| {
+            trust.workspace_id == self.workspace_id
+                && trust.backend_url == api.base_url
+                && trust.state == WorkspaceIssuerTrustState::Active
+        }) {
+            return Err(reject("request does not match initialized Workspace trust"));
+        }
+        let stage = options
+            .root
+            .join(format!(".fixture-{}", uuid::Uuid::now_v7()));
+        io(fs::create_dir(&stage))?;
+        let store = FsRuntimeStore {
+            root: stage.clone(),
+            _owner_lock: Some(lock),
+            worker_write_fault: Arc::new(AtomicU8::new(AtomicWriteFault::None as u8)),
+            #[cfg(test)]
+            worker_reconcile_sync_fault: Default::default(),
+            #[cfg(test)]
+            worker_reconcile_read_fault: Default::default(),
+            #[cfg(test)]
+            worker_reconcile_counts: Default::default(),
+        };
+        let result = (|| {
+            let worker = PersistedWorkerRecord {
+                restore_guard: crate::runtime::RestoreGuard::default(),
+                worker_ref: WorkerRef::new(self.worker_id),
+                worker_id: self.worker_id,
+                profile: request.profile.clone(),
+                display_name: request.display_name.clone(),
+                profile_source: request.profile_source.reference(),
+                config_bundle: request.config_bundle.clone(),
+                created_at_ms: self.created_at_ms,
+                status: WorkerStatus::Stopped,
+                last_finished_submission_request_id: None,
+                workspace_id: Some(self.workspace_id.clone()),
+                workdir_attachments: Vec::new(),
+                logical_workdir_attachments: Vec::new(),
+                execution_state: PersistedWorkerExecutionState::Available(
+                    PersistedWorkerExecution {
+                        request,
+                        binding: None,
+                        restore_intent: WorkerRestoreIntent::Explicit,
+                    },
+                ),
+            };
+            WorkerAggregateRecord::from_persisted(&worker)
+                .validate(&store.worker_dir(&self.worker_id).join(WORKER_FILE))
+                .map_err(|_| reject("invalid installation request"))?;
+            store.write_worker_record(&worker)?;
+            let worker_dir = store.worker_dir(&self.worker_id);
+            atomic_write_json(
+                &worker_dir.join(WORKER_METADATA_FILE),
+                &metadata,
+                "write fixture metadata",
+            )?;
+            let session = worker_dir.join("session");
+            io(fs::create_dir_all(session.join("segments")))?;
+            write_new(&session.join("session.json"), &self.session_manifest)?;
+            write_new(
+                &session
+                    .join("segments")
+                    .join(format!("{}.jsonl", self.segment_id)),
+                &self.segment_jsonl,
+            )?;
+            self.validate_session(&session)?;
+            store.write_runtime_snapshot(&PersistedRuntimeState {
+                subject_create_recovery_fences: BTreeMap::new(),
+                subject_releases: Vec::new(),
+                // Match normal store initialization: this is catalog availability,
+                // not a process or Worker execution start.
+                display_name: options.display_name.clone(),
+                status: RuntimeStatus::Running,
+                next_diagnostic_id: 1,
+                workers: BTreeMap::new(),
+                workspace_owners: BTreeMap::new(),
+                diagnostics: Vec::new(),
+            })?;
+            let loaded = store.load_runtime_state()?;
+            if loaded.workers.len() != 1 || !loaded.diagnostics.is_empty() {
+                return Err(reject("staged aggregate failed ordinary store roundtrip"));
+            }
+            sync_runtime_tree(&stage)?;
+            // No destination existed when checked under the owner lock. Hard-link
+            // the snapshot last with create-new semantics; never replace a file.
+            // Reserve the destination catalog with create-new semantics. Even
+            // an empty preexisting workers directory must never be replaced.
+            let destination_workers = options.root.join(WORKERS_DIR);
+            io(fs::create_dir(&destination_workers))?;
+            io(fs::rename(
+                store.worker_dir(&self.worker_id),
+                destination_workers.join(self.worker_id.to_string()),
+            ))?;
+            sync_directory(&destination_workers, "publish stopped Worker aggregate")?;
+            io(fs::hard_link(
+                stage.join(RUNTIME_FILE),
+                options.root.join(RUNTIME_FILE),
+            ))?;
+            sync_directory(&options.root, "publish stopped fixture")?;
+            Ok(())
+        })();
+        // Only remove our private staging directory, never existing destination data.
+        let cleanup = fs::remove_dir_all(&stage);
+        result?;
+        io(cleanup)?;
+        sync_directory(&options.root, "finish stopped fixture")
+    }
+
+    fn validate_shape(&self) -> Result<(), RuntimeError> {
+        // Version 1 templates carry only settled Session material, never a
+        // lifecycle operation. Schema 12 -> 13 changes pending decision evidence,
+        // not that payload. Accept the fixed schema-12 artifact read-only and
+        // install through the current aggregate owner; never rewrite its bytes.
+        if self.template_version != TEMPLATE_VERSION
+            || !matches!(
+                self.runtime_schema_version,
+                RESTORE_ADMISSION_SCHEMA_VERSION | SCHEMA_VERSION
+            )
+        {
+            return Err(reject("unsupported template schema"));
+        }
+        if !portable_id(&self.runtime_id)
+            || !portable_id(&self.workspace_id)
+            || self.session_id.get_version_num() != 7
+            || self.segment_id.get_version_num() != 7
+            || self.session_manifest.len() > MAX_SESSION_BYTES
+            || self.segment_jsonl.len() > MAX_SESSION_BYTES
+        {
+            return Err(reject("invalid portable identity or size"));
+        }
+        let manifest: SessionManifest = decode(&self.session_manifest)?;
+        if manifest.session_id != self.session_id
+            || manifest.schema_version == 0
+            || !self.segment_jsonl.ends_with(b"\n")
+        {
+            return Err(reject("Session identity or committed boundary mismatch"));
+        }
+        // Read-only JSON traversal, never mutation/reconstruction of Session bytes.
+        // Reject known secret/host markers even in otherwise valid typed records.
+        for line in self
+            .segment_jsonl
+            .split(|b| *b == b'\n')
+            .filter(|line| !line.is_empty())
+        {
+            let _: PortableJson = decode(line)?;
+            validate_preserved_record(line)?;
+        }
+        Ok(())
+    }
+
+    fn validate_history_entry(
+        &self,
+        entry: &session_store::LoggedHistoryEntry,
+        ids: &mut std::collections::BTreeSet<String>,
+    ) -> Result<(), RuntimeError> {
+        if !portable_id(&entry.metadata.entry_id.0)
+            || !ids.insert(entry.metadata.entry_id.0.clone())
+        {
+            return Err(reject("duplicate or invalid entry identity"));
+        }
+        if entry
+            .metadata
+            .input_metadata
+            .as_ref()
+            .is_some_and(|metadata| {
+                metadata.feature_payload().is_some() || metadata.public_item().is_some()
+            })
+        {
+            // These are opaque owner-specific payloads, not part of the narrow
+            // artificial conversation contract. Never preserve unchecked JSON.
+            return Err(reject("Feature history payloads are unsupported"));
+        }
+        if !matches!(
+            entry.item,
+            session_store::LoggedItem::Message { .. }
+                | session_store::LoggedItem::AssistantResponseBoundary { .. }
+        ) {
+            return Err(reject("only text conversation records are supported"));
+        }
+        if let session_store::LoggedSessionHistoryOrigin::ModelOutput { worker } =
+            &entry.metadata.origin
+        {
+            // Session producer IDs identify the original execution, not the
+            // Workspace catalog Worker. Preserve them; do not rebind/rewrite them.
+            if !portable_id(&worker.worker_id)
+                || worker
+                    .workspace_id
+                    .as_ref()
+                    .is_some_and(|id| id != &self.workspace_id)
+                || worker
+                    .runtime_id
+                    .as_ref()
+                    .is_some_and(|id| id != &self.runtime_id)
+            {
+                return Err(reject("Session origin identity mismatch"));
+            }
+        }
+        Ok(())
+    }
+
+    fn validate_session(&self, root: &Path) -> Result<(), RuntimeError> {
+        for entry in io(fs::read_dir(root))? {
+            let entry = io(entry)?;
+            match entry.file_name().to_str() {
+                Some("session.json" | "segments") => {}
+                Some("artifacts") => check_empty_artifacts(&entry.path())?,
+                _ => return Err(reject("unsupported Session files")),
+            }
+        }
+        let entries = io(fs::read_dir(root.join("segments")))?.collect::<Result<Vec<_>, _>>();
+        let entries = io(entries)?;
+        if entries.len() != 1
+            || entries[0].file_name() != format!("{}.jsonl", self.segment_id).as_str()
+        {
+            return Err(reject("only the fixed single-Segment fixture is supported"));
+        }
+        let store = WorkerSessionStore::open_read_only(root)
+            .map_err(|_| reject("unsupported Session schema"))?;
+        let logs = store
+            .read_all_read_only_bounded(self.session_id, self.segment_id, MAX_SESSION_BYTES as u64)
+            .map_err(|_| reject("invalid canonical Session records"))?
+            .0;
+        if !matches!(logs.first(), Some(session_store::LogEntry::AnnotatedSegmentStart {
+            session_id, system_prompt: Some(_), forked_from: None, compacted_from: None, ..
+        }) if *session_id == self.session_id)
+        {
+            return Err(reject(
+                "missing system prompt, invalid Segment seed, or unsupported lineage",
+            ));
+        }
+        let mut ids = std::collections::BTreeSet::new();
+        for (index, log) in logs.iter().enumerate() {
+            use session_store::LogEntry;
+            match log {
+                LogEntry::AnnotatedSegmentStart { history, .. }
+                    if index == 0 && history.is_empty() => {}
+                LogEntry::AnnotatedUserInput {
+                    history,
+                    extensions,
+                    ..
+                } => {
+                    for extension in extensions {
+                        validate_settled_activation(extension)?;
+                    }
+                    for entry in history {
+                        self.validate_history_entry(entry, &mut ids)?;
+                    }
+                }
+                LogEntry::AnnotatedAssistantItem { entry, .. } => {
+                    self.validate_history_entry(entry, &mut ids)?
+                }
+                LogEntry::RunCompleted {
+                    entry_id: Some(id), ..
+                } => {
+                    if !portable_id(&id.0) || !ids.insert(id.0.clone()) {
+                        return Err(reject("duplicate or invalid entry identity"));
+                    }
+                }
+                LogEntry::Invoke { .. } | LogEntry::LlmUsage { .. } | LogEntry::TurnEnd { .. } => {}
+                _ => {
+                    return Err(reject(
+                        "unsupported conversation record or pending operation",
+                    ));
+                }
+            }
+        }
+        let state = session_store::collect_state(&logs);
+        if state.last_run_interrupted || state.active_run_turn_count.is_some() {
+            return Err(reject("Session has pending execution"));
+        }
+        for segments in &state.user_segments {
+            if segments
+                .iter()
+                .any(|segment| !matches!(segment, protocol::Segment::Text { .. }))
+            {
+                return Err(reject("Session attachments are unsupported"));
+            }
+        }
+        Ok(())
+    }
+}
+
+// Compare *fields*, not re-encoded bytes or scalar values (e.g. f32 JSON
+// normalization). Serde's permissive/untagged/flattened compatibility readers
+// can drop unknown data without surfacing it to an ignored-field callback. A
+// retained field must survive the owner's typed projection at every depth.
+// This intentionally accepts only canonical writer shapes: even an explicit
+// empty optional field omitted by that writer is rejected, not repaired.
+// The projection is validation-only; installation always writes original bytes.
+fn validate_preserved_record(bytes: &[u8]) -> Result<(), RuntimeError> {
+    let original: serde_json::Value = decode(bytes)?;
+    let record: session_store::LogEntry = decode(bytes)?;
+    let projection =
+        serde_json::to_value(&record).map_err(|_| reject("invalid typed Session record"))?;
+    require_preserved_fields(&original, &projection)
+}
+
+fn require_preserved_fields(
+    original: &serde_json::Value,
+    projection: &serde_json::Value,
+) -> Result<(), RuntimeError> {
+    use serde_json::Value;
+    match (original, projection) {
+        (Value::Object(original), Value::Object(projection)) => {
+            for (key, value) in original {
+                let projected = projection
+                    .get(key)
+                    .ok_or_else(|| reject("unknown or noncanonical Session field"))?;
+                require_preserved_fields(value, projected)?;
+            }
+        }
+        (Value::Array(original), Value::Array(projection))
+            if original.len() == projection.len() =>
+        {
+            for (value, projected) in original.iter().zip(projection) {
+                require_preserved_fields(value, projected)?;
+            }
+        }
+        (Value::Null, Value::Null)
+        | (Value::Bool(_), Value::Bool(_))
+        | (Value::Number(_), Value::Number(_))
+        | (Value::String(_), Value::String(_)) => {}
+        _ => return Err(reject("noncanonical Session field shape")),
+    }
+    Ok(())
+}
+
+// Exact receipt wire fields owned by worker::PendingActivationState. Keep these
+// fixture-only mirrors strict; do not change the production compatibility reader.
+#[derive(Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct SubmissionReceipt {
+    submission_request_id: String,
+    source_namespace: String,
+    submission_id: String,
+    payload_digest: String,
+    disposition: protocol::SubmissionDisposition,
+}
+
+#[derive(Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct NotificationReceipt {
+    notification_request_id: String,
+    source_namespace: String,
+    payload_digest: String,
+}
+
+// The approved conversation has committed submission checkpoints in its original
+// Session records. Keep their receipt bytes, but reject any resumable queue or
+// Feature invocation history. This projection confers no execution authority.
+fn validate_settled_activation(
+    extension: &session_store::SessionExtension,
+) -> Result<(), RuntimeError> {
+    #[derive(Deserialize)]
+    #[serde(deny_unknown_fields)]
+    struct Checkpoint {
+        #[serde(rename = "next_activation_sequence")]
+        _next_activation_sequence: u64,
+        activating: Option<serde::de::IgnoredAny>,
+        activating_notification: Vec<serde::de::IgnoredAny>,
+        pending: Vec<serde::de::IgnoredAny>,
+        pending_notifications: Vec<serde::de::IgnoredAny>,
+        #[serde(rename = "receipts")]
+        _receipts: Vec<SubmissionReceipt>,
+        #[serde(rename = "notification_receipts")]
+        _notification_receipts: Vec<NotificationReceipt>,
+        invocation_receipts: Vec<serde::de::IgnoredAny>,
+        invocation_recovery_blocked: bool,
+    }
+    if extension.domain != "worker.pending_activations.v1" {
+        return Err(reject("unsupported Session extension"));
+    }
+    let checkpoint: Checkpoint = serde_json::from_value(extension.payload.clone())
+        .map_err(|_| reject("invalid Session activation checkpoint"))?;
+    if checkpoint.activating.is_some()
+        || !checkpoint.activating_notification.is_empty()
+        || !checkpoint.pending.is_empty()
+        || !checkpoint.pending_notifications.is_empty()
+        || !checkpoint.invocation_receipts.is_empty()
+        || checkpoint.invocation_recovery_blocked
+    {
+        return Err(reject(
+            "Session has pending activation or invocation history",
+        ));
+    }
+    Ok(())
+}
+
+fn check_request(request: &CreateWorkerRequest, install: bool) -> Result<(), RuntimeError> {
+    if !request.workdir_attachment_requests.is_empty()
+        || !request.workdir_attachments.is_empty()
+        || request.backend_job.is_some()
+        || request.subjektiv_attached
+        || request.worker_observation_enabled
+        || !request.worker_observation_grants.is_empty()
+        || (install && request.initial_input.is_some())
+    {
+        return Err(reject(
+            "unsupported capabilities, attachments, or initial input",
+        ));
+    }
+    if install
+        && (request.create_fingerprint.trim().is_empty()
+            || request.config_bundle.is_none()
+            || !matches!(
+                request.profile_source,
+                ProfileSourceArchiveSource::WorkspaceConfig { .. }
+            ))
+    {
+        return Err(reject(
+            "fresh config/profile references and create fingerprint are required",
+        ));
+    }
+    Ok(())
+}
+
+fn check_empty_artifacts(root: &Path) -> Result<(), RuntimeError> {
+    for entry in io(fs::read_dir(root))? {
+        let entry = io(entry)?;
+        let metadata = io(entry.metadata())?;
+        if metadata.is_dir() {
+            check_empty_artifacts(&entry.path())?;
+        } else if entry.file_name() != ".aggregate.lock" || metadata.len() != 0 {
+            return Err(reject("Session artifacts are unsupported"));
+        }
+    }
+    Ok(())
+}
+
+fn write_new(path: &Path, bytes: &[u8]) -> Result<(), RuntimeError> {
+    let mut file = io(OpenOptions::new().write(true).create_new(true).open(path))?;
+    io(file.write_all(bytes))?;
+    io(file.sync_all())
+}
+
+// A streaming validation visitor, not a generic document available to callers.
+struct PortableJson;
+impl<'de> Deserialize<'de> for PortableJson {
+    fn deserialize<D: serde::Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        struct Visitor;
+        impl<'de> serde::de::Visitor<'de> for Visitor {
+            type Value = PortableJson;
+            fn expecting(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+                f.write_str("portable fixture JSON")
+            }
+            fn visit_bool<E: serde::de::Error>(self, _: bool) -> Result<PortableJson, E> {
+                Ok(PortableJson)
+            }
+            fn visit_i64<E: serde::de::Error>(self, _: i64) -> Result<PortableJson, E> {
+                Ok(PortableJson)
+            }
+            fn visit_u64<E: serde::de::Error>(self, _: u64) -> Result<PortableJson, E> {
+                Ok(PortableJson)
+            }
+            fn visit_f64<E: serde::de::Error>(self, _: f64) -> Result<PortableJson, E> {
+                Ok(PortableJson)
+            }
+            fn visit_unit<E: serde::de::Error>(self) -> Result<PortableJson, E> {
+                Ok(PortableJson)
+            }
+            fn visit_str<E: serde::de::Error>(self, value: &str) -> Result<PortableJson, E> {
+                let lower = value.to_ascii_lowercase();
+                if [
+                    "://",
+                    "private key",
+                    "bearer ",
+                    "sk-",
+                    "api_key",
+                    "access_token",
+                    "refresh_token",
+                    "private_key",
+                    "authorization",
+                ]
+                .iter()
+                .any(|marker| lower.contains(marker))
+                    || value
+                        .split_whitespace()
+                        .any(|word| word.starts_with('/') || word.contains(":\\"))
+                {
+                    return Err(E::custom("nonportable or sensitive fixture material"));
+                }
+                Ok(PortableJson)
+            }
+            fn visit_seq<A: serde::de::SeqAccess<'de>>(
+                self,
+                mut seq: A,
+            ) -> Result<PortableJson, A::Error> {
+                while seq.next_element::<PortableJson>()?.is_some() {}
+                Ok(PortableJson)
+            }
+            fn visit_map<A: serde::de::MapAccess<'de>>(
+                self,
+                mut map: A,
+            ) -> Result<PortableJson, A::Error> {
+                let mut keys = std::collections::BTreeSet::new();
+                while let Some(key) = map.next_key::<String>()? {
+                    // A duplicate key can hide discarded data even inside an
+                    // otherwise opaque extension payload. Never retain it.
+                    if !keys.insert(key.clone()) {
+                        return Err(serde::de::Error::custom("duplicate fixture JSON field"));
+                    }
+                    serde::de::Visitor::visit_str::<A::Error>(Visitor, &key)?;
+                    map.next_value::<PortableJson>()?;
+                }
+                Ok(PortableJson)
+            }
+        }
+        deserializer.deserialize_any(Visitor)
+    }
+}

crates/worker-runtime/src/test_support_tests.rs

@@ -0,0 +1,919 @@
+use super::*;
+use crate::auth::RuntimeIdentityMaterial;
+use crate::catalog::{ProfileSelector, WorkspaceApiRef};
+use crate::profile_archive::ProfileSourceGraphSummary;
+use session_store::{LogEntry, Store};
+use sha2::{Digest, Sha256};
+
+const RUNTIME: &str = "fixture-runtime";
+const WORKSPACE: &str = "fixture-workspace";
+const NEW_URL: &str = "http://127.0.0.1:34567";
+
+#[derive(Serialize)]
+#[serde(untagged)]
+enum ReceiptFixture<T> {
+    Valid(T),
+    Malformed(u8),
+}
+
+#[derive(Default, Serialize)]
+struct CheckpointFixture {
+    next_activation_sequence: u64,
+    activating: Option<u8>,
+    activating_notification: Vec<u8>,
+    pending: Vec<u8>,
+    pending_notifications: Vec<u8>,
+    receipts: Vec<ReceiptFixture<SubmissionReceipt>>,
+    notification_receipts: Vec<ReceiptFixture<NotificationReceipt>>,
+    invocation_receipts: Vec<u8>,
+    invocation_recovery_blocked: bool,
+}
+
+fn settled_checkpoint() -> CheckpointFixture {
+    CheckpointFixture {
+        next_activation_sequence: 2,
+        receipts: vec![ReceiptFixture::Valid(SubmissionReceipt {
+            submission_request_id: "completed-submission-request".into(),
+            source_namespace: "fixture-source".into(),
+            submission_id: "completed-submission".into(),
+            payload_digest: "fixture-payload-digest".into(),
+            disposition: protocol::SubmissionDisposition::Started,
+        })],
+        notification_receipts: vec![ReceiptFixture::Valid(NotificationReceipt {
+            notification_request_id: "completed-notification-request".into(),
+            source_namespace: "fixture-source".into(),
+            payload_digest: "fixture-notification-digest".into(),
+        })],
+        ..Default::default()
+    }
+}
+
+fn checkpoint_extension(checkpoint: &CheckpointFixture) -> session_store::SessionExtension {
+    session_store::SessionExtension::new(
+        "worker.pending_activations.v1",
+        serde_json::to_value(checkpoint).unwrap(),
+    )
+}
+
+fn worker_id() -> WorkerId {
+    "01a1657d-b762-7791-87ae-a49f28cadbce".parse().unwrap()
+}
+
+fn resolved() -> manifest::WorkerManifest {
+    let mut config = manifest::WorkerManifestConfig::resolution_defaults();
+    config.worker.name = Some(runtime_worker_name(worker_id()));
+    manifest::WorkerManifest::try_from(config).unwrap()
+}
+
+fn request() -> CreateWorkerRequest {
+    CreateWorkerRequest {
+        worker_id: worker_id(),
+        create_fingerprint: "new-create-fingerprint".into(),
+        profile: ProfileSelector::Builtin("default".into()),
+        display_name: Some("fixture".into()),
+        profile_source: ProfileSourceArchiveSource::WorkspaceConfig {
+            archive: ProfileSourceArchiveRef {
+                id: "new-profile".into(),
+                digest: "new-profile-digest".into(),
+                size_bytes: 1,
+                source_graph: ProfileSourceGraphSummary {
+                    source_count: 1,
+                    total_source_bytes: 1,
+                    entrypoints: BTreeMap::from([("builtin:default".into(), "default.dcl".into())]),
+                    import_count: 0,
+                },
+            },
+        },
+        config_bundle: Some(ConfigBundleRef {
+            id: "new-config".into(),
+            digest: "new-config-digest".into(),
+        }),
+        workspace_api: Some(WorkspaceApiRef {
+            workspace_id: WORKSPACE.into(),
+            base_url: NEW_URL.into(),
+        }),
+        initial_input: None,
+        workdir_attachment_requests: Vec::new(),
+        workdir_attachments: Vec::new(),
+        worker_observation_enabled: false,
+        worker_observation_grants: Vec::new(),
+        memory_settings: Some(manifest::WorkspaceMemorySettingsSnapshot {
+            workspace_id: WORKSPACE.into(),
+            language: "en".into(),
+        }),
+        subjektiv_attached: false,
+        backend_job: None,
+    }
+}
+
+fn initialize(root: &Path, runtime_id: &str, workspace_id: &str, url: &str) {
+    #[derive(Serialize)]
+    struct Auth {
+        identity: RuntimeIdentityMaterial,
+        workspace_issuers: Vec<WorkspaceIssuerTrustRecord>,
+    }
+    let issuer = RuntimeIdentityMaterial::generate("fixture-issuer").unwrap();
+    let digest = Sha256::digest(crate::auth::decode_public_key(&issuer.public_key).unwrap());
+    let fingerprint = format!(
+        "sha256:{}",
+        digest
+            .iter()
+            .map(|byte| format!("{byte:02x}"))
+            .collect::<String>()
+    );
+    let trust = WorkspaceIssuerTrustRecord::from_bundle(
+        server_api::WorkspacePublicIdentityBundle {
+            workspace_id: workspace_id.into(),
+            backend_url: url.into(),
+            key_id: "fixture-issuer".into(),
+            algorithm: "ed25519".into(),
+            public_key: issuer.public_key,
+            public_key_fingerprint: fingerprint,
+        },
+        "fixture-trust".into(),
+        100,
+    )
+    .unwrap();
+    fs::create_dir_all(root).unwrap();
+    fs::write(
+        root.join("auth.toml"),
+        toml::to_string(&Auth {
+            identity: RuntimeIdentityMaterial::generate(runtime_id).unwrap(),
+            workspace_issuers: vec![trust],
+        })
+        .unwrap(),
+    )
+    .unwrap();
+}
+
+fn record(request: CreateWorkerRequest) -> PersistedWorkerRecord {
+    PersistedWorkerRecord {
+        restore_guard: crate::runtime::RestoreGuard::default(),
+        worker_ref: WorkerRef::new(request.worker_id),
+        worker_id: request.worker_id,
+        profile: request.profile.clone(),
+        display_name: request.display_name.clone(),
+        profile_source: request.profile_source.reference(),
+        config_bundle: request.config_bundle.clone(),
+        created_at_ms: Some(123456),
+        status: WorkerStatus::Stopped,
+        last_finished_submission_request_id: Some("old-submission-operation".into()),
+        workspace_id: Some(WORKSPACE.into()),
+        workdir_attachments: vec![],
+        logical_workdir_attachments: vec![],
+        execution_state: PersistedWorkerExecutionState::Available(PersistedWorkerExecution {
+            request,
+            binding: Some(PersistedWorkerExecutionBinding {}),
+            restore_intent: WorkerRestoreIntent::Explicit,
+        }),
+    }
+}
+
+fn source() -> (tempfile::TempDir, PathBuf, StoppedWorkerTemplate) {
+    let temp = tempfile::tempdir().unwrap();
+    let root = temp.path().join("runtime");
+    initialize(&root, RUNTIME, WORKSPACE, "http://127.0.0.1:23456");
+    drop(
+        crate::Runtime::with_fs_store(FsRuntimeStoreOptions::new(&root).with_runtime_id(RUNTIME))
+            .unwrap(),
+    );
+    let opened = FsRuntimeStore::open_or_create(root.clone(), RUNTIME).unwrap();
+    let mut old = request();
+    old.workspace_api.as_mut().unwrap().base_url = "http://127.0.0.1:23456".into();
+    old.config_bundle.as_mut().unwrap().id = "old-config".into();
+    opened.store.write_worker_record(&record(old)).unwrap();
+    let worker = opened.store.worker_dir(&worker_id());
+    let session_id = "01a1657d-b800-7777-8888-111111111111".parse().unwrap();
+    let segment_id = "01a1657d-b818-77c1-bd73-176172a4bd82".parse().unwrap();
+    let mut metadata = WorkerMetadata::new(
+        runtime_worker_name(worker_id()),
+        Some(WorkerActiveSegmentRef {
+            session_id,
+            segment_id: Some(segment_id),
+        }),
+    )
+    .with_workspace_id(WORKSPACE)
+    .with_workspace_root(PathBuf::from("/old/host/path"));
+    metadata.resolved_manifest_snapshot =
+        Some(manifest::write_persisted_worker_manifest_snapshot(&resolved()).unwrap());
+    atomic_write_json(
+        &worker.join(WORKER_METADATA_FILE),
+        &metadata,
+        "test metadata",
+    )
+    .unwrap();
+    let sessions = WorkerSessionStore::new(worker.join("session")).unwrap();
+    let history = session_store::LoggedHistoryEntry {
+        item: session_store::LoggedItem::Message {
+            role: session_store::LoggedRole::Assistant,
+            content: vec![session_store::LoggedContentPart::Text {
+                text: "artificial fixture reply".into(),
+            }],
+        },
+        metadata: session_store::LoggedSessionHistoryMetadata {
+            entry_id: session_store::LoggedSessionHistoryEntryId("fixed-entry-id".into()),
+            origin: session_store::LoggedSessionHistoryOrigin::ModelOutput {
+                worker: session_store::LoggedWorkerSubject {
+                    workspace_id: Some(WORKSPACE.into()),
+                    runtime_id: Some(RUNTIME.into()),
+                    worker_id: worker_id().to_string(),
+                },
+            },
+            derivation: None,
+            input_metadata: None,
+        },
+    };
+    sessions
+        .create_segment(
+            session_id,
+            segment_id,
+            &[
+                LogEntry::AnnotatedSegmentStart {
+                    ts: 123,
+                    session_id,
+                    system_prompt: Some("Harmless artificial fixture instructions.".into()),
+                    config: Default::default(),
+                    history: vec![],
+                    forked_from: None,
+                    compacted_from: None,
+                },
+                LogEntry::Invoke {
+                    ts: 124,
+                    trigger: protocol::InvokeKind::UserSend,
+                },
+                LogEntry::AnnotatedUserInput {
+                    ts: 124,
+                    segments: vec![protocol::Segment::text(
+                        "Harmless artificial fixture input.",
+                    )],
+                    extensions: vec![checkpoint_extension(&settled_checkpoint())],
+                    history: vec![session_store::LoggedHistoryEntry {
+                        item: session_store::LoggedItem::Message {
+                            role: session_store::LoggedRole::User,
+                            content: vec![session_store::LoggedContentPart::Text {
+                                text: "Harmless artificial fixture input.".into(),
+                            }],
+                        },
+                        metadata: session_store::LoggedSessionHistoryMetadata {
+                            entry_id: session_store::LoggedSessionHistoryEntryId(
+                                "fixed-user-id".into(),
+                            ),
+                            origin: session_store::LoggedSessionHistoryOrigin::BackendInstruction {
+                                operation_id: None,
+                            },
+                            derivation: None,
+                            input_metadata: None,
+                        },
+                    }],
+                },
+                LogEntry::AnnotatedAssistantItem {
+                    ts: 125,
+                    entry: history,
+                },
+                LogEntry::LlmUsage {
+                    ts: 125,
+                    history_len: 2,
+                    input_total_tokens: 1,
+                    cache_read_tokens: 0,
+                    cache_write_tokens: 0,
+                    output_tokens: 1,
+                },
+                LogEntry::TurnEnd {
+                    ts: 126,
+                    turn_count: 1,
+                },
+                LogEntry::RunCompleted {
+                    entry_id: Some(session_store::LoggedSessionHistoryEntryId(
+                        "fixed-terminal-id".into(),
+                    )),
+                    ts: 126,
+                    interrupted: false,
+                    result: agen::EngineResult::Finished,
+                    active_run_turn_count: None,
+                },
+            ],
+        )
+        .unwrap();
+    fs::create_dir_all(worker.join("runs/old-run")).unwrap();
+    fs::write(worker.join("runs/old-run/private.log"), "not exported").unwrap();
+    drop(opened);
+    let template = StoppedWorkerTemplate::extract(&worker).unwrap();
+    (temp, worker, template)
+}
+
+fn destination(template: &StoppedWorkerTemplate) -> (tempfile::TempDir, FsRuntimeStoreOptions) {
+    let temp = tempfile::tempdir().unwrap();
+    let root = temp.path().join("runtime");
+    initialize(
+        &root,
+        template.runtime_id(),
+        template.workspace_id(),
+        NEW_URL,
+    );
+    (
+        temp,
+        FsRuntimeStoreOptions::new(root).with_runtime_id(template.runtime_id()),
+    )
+}
+
+fn assert_fresh(options: &FsRuntimeStoreOptions) {
+    assert_eq!(fs::read_dir(&options.root).unwrap().count(), 1);
+    assert!(options.root.join("auth.toml").is_file());
+}
+
+#[test]
+fn extraction_install_roundtrip_preserves_bytes_and_is_inactive() {
+    let (_source, worker, template) = source();
+    let original_worker = fs::read(worker.join(WORKER_FILE)).unwrap();
+    let original_metadata = fs::read(worker.join(WORKER_METADATA_FILE)).unwrap();
+    let exported = serde_json::to_vec(&template).unwrap();
+    let exported_text = String::from_utf8(exported.clone()).unwrap();
+    for forbidden in [
+        "old-config",
+        "old-submission-operation",
+        "restore_guard",
+        "private_key",
+        "http://",
+        "/old/host",
+        "private.log",
+        "resolved_manifest_snapshot",
+    ] {
+        assert!(!exported_text.contains(forbidden));
+    }
+    let template: StoppedWorkerTemplate = serde_json::from_slice(&exported).unwrap();
+    let (_target, options) = destination(&template);
+    let auth_before = fs::read(options.root.join("auth.toml")).unwrap();
+    template.install(&options, request(), &resolved()).unwrap();
+    assert_eq!(
+        fs::read(options.root.join("auth.toml")).unwrap(),
+        auth_before
+    );
+    assert_eq!(fs::read(worker.join(WORKER_FILE)).unwrap(), original_worker);
+    assert_eq!(
+        fs::read(worker.join(WORKER_METADATA_FILE)).unwrap(),
+        original_metadata
+    );
+    let installed = options
+        .root
+        .join(WORKERS_DIR)
+        .join(template.worker_id().to_string());
+    assert_eq!(
+        fs::read(installed.join("session/session.json")).unwrap(),
+        template.session_manifest
+    );
+    assert_eq!(
+        fs::read(installed.join(format!("session/segments/{}.jsonl", template.segment_id())))
+            .unwrap(),
+        template.segment_jsonl
+    );
+    assert!(!installed.join("runs").exists());
+    let metadata: WorkerMetadata =
+        decode(&fs::read(installed.join(WORKER_METADATA_FILE)).unwrap()).unwrap();
+    assert_eq!(metadata.active.unwrap().session_id, template.session_id());
+    assert!(metadata.workspace_root.is_none() && metadata.resolved_manifest_snapshot.is_some());
+
+    let opened = FsRuntimeStore::open_or_create(options.root.clone(), RUNTIME).unwrap();
+    let state = opened.state.unwrap();
+    assert!(state.diagnostics.is_empty());
+    let saved = state.workers.get(&worker_id()).unwrap();
+    assert_eq!(saved.status, WorkerStatus::Stopped);
+    assert!(saved.last_finished_submission_request_id.is_none());
+    let PersistedWorkerExecutionState::Available(execution) = &saved.execution_state else {
+        panic!("not available")
+    };
+    assert_eq!(execution.restore_intent, WorkerRestoreIntent::Explicit);
+    assert!(execution.binding.is_none());
+    assert_eq!(execution.request, request());
+    // Fresh restore guard, no operation owner/receipt history.
+    assert_ne!(
+        serde_json::to_vec(&saved.restore_guard).unwrap(),
+        serde_json::to_vec(
+            &decode::<WorkerAggregateRecord>(&original_worker)
+                .unwrap()
+                .restore_guard
+        )
+        .unwrap()
+    );
+    let guard: GuardState = decode(&serde_json::to_vec(&saved.restore_guard).unwrap()).unwrap();
+    assert!(guard.active_request_id.is_none() && guard.owners.is_empty());
+    drop(opened.store);
+
+    struct NoExecution;
+    impl crate::execution::WorkerExecutionBackend for NoExecution {
+        fn backend_id(&self) -> &str {
+            "fixture-no-execution"
+        }
+        fn spawn_worker(
+            &self,
+            _: crate::execution::WorkerExecutionSpawnRequest,
+        ) -> crate::execution::WorkerExecutionSpawnResult {
+            panic!("fixture executed")
+        }
+        fn dispatch_input(
+            &self,
+            _: &WorkerRef,
+            _: crate::interaction::WorkerInput,
+        ) -> crate::execution::WorkerExecutionResult {
+            panic!("fixture submitted")
+        }
+        fn restore_worker(
+            &self,
+            _: crate::execution::WorkerExecutionRestoreRequest,
+        ) -> crate::execution::WorkerExecutionSpawnResult {
+            panic!("fixture auto-restored")
+        }
+    }
+    let runtime =
+        crate::Runtime::with_fs_store_and_execution_backend(options.clone(), Arc::new(NoExecution))
+            .unwrap();
+    assert_eq!(runtime.list_workers().unwrap().len(), 1);
+    assert_eq!(runtime.summary().unwrap().active_worker_count, 0);
+    assert_eq!(
+        runtime
+            .worker_detail(&WorkerRef::new(worker_id()))
+            .unwrap()
+            .status,
+        WorkerStatus::Stopped
+    );
+    drop(runtime);
+    let runtime = crate::Runtime::with_fs_store(options).unwrap();
+    assert_eq!(
+        runtime.list_workers().unwrap()[0].status,
+        WorkerStatus::Stopped
+    );
+}
+
+#[test]
+fn install_rejects_schema_identity_input_and_overwrite() {
+    let (_source, _, template) = source();
+    for change in [
+        "template",
+        "runtime-schema",
+        "session-schema",
+        "session-id",
+        "segment-id",
+        "runtime-id",
+        "workspace-id",
+        "worker-id",
+        "origin-id",
+        "duplicate-id",
+        "truncated",
+    ] {
+        let (_target, options) = destination(&template);
+        let mut invalid = template.clone();
+        match change {
+            "template" => invalid.template_version += 1,
+            "runtime-schema" => invalid.runtime_schema_version += 1,
+            "session-schema" => {
+                invalid.session_manifest = format!(
+                    "{{\"schema_version\":999,\"session_id\":\"{}\"}}",
+                    invalid.session_id
+                )
+                .into_bytes()
+            }
+            "session-id" => invalid.session_id = uuid::Uuid::now_v7(),
+            "segment-id" => invalid.segment_id = uuid::Uuid::nil(), // no external attestation: valid new Segment ID is otherwise portable
+            "runtime-id" => invalid.runtime_id = "other-runtime".into(),
+            "workspace-id" => invalid.workspace_id = "other-workspace".into(),
+            "worker-id" => invalid.worker_id = WorkerId::now_v7(),
+            "origin-id" => {
+                invalid.segment_jsonl = String::from_utf8(invalid.segment_jsonl)
+                    .unwrap()
+                    .replace(RUNTIME, "other-runtime")
+                    .into_bytes()
+            }
+            "duplicate-id" => {
+                let last = invalid
+                    .segment_jsonl
+                    .split_inclusive(|b| *b == b'\n')
+                    .last()
+                    .unwrap()
+                    .to_vec();
+                invalid.segment_jsonl.extend(last);
+            }
+            "truncated" => {
+                invalid.segment_jsonl.pop();
+            }
+            _ => unreachable!(),
+        }
+        assert!(
+            invalid.install(&options, request(), &resolved()).is_err(),
+            "accepted {change}"
+        );
+        assert_fresh(&options);
+    }
+    let (_target, options) = destination(&template);
+    let mut launch = request();
+    launch.initial_input = Some(crate::interaction::WorkerInput::user("do not execute"));
+    assert!(template.install(&options, launch, &resolved()).is_err());
+    assert_fresh(&options);
+    template.install(&options, request(), &resolved()).unwrap();
+    let worker_path = options
+        .root
+        .join(WORKERS_DIR)
+        .join(worker_id().to_string())
+        .join(WORKER_FILE);
+    let before = fs::read(&worker_path).unwrap();
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    assert_eq!(fs::read(worker_path).unwrap(), before);
+}
+
+#[test]
+fn rejects_uninitialized_untrusted_live_and_partial_destination() {
+    let (_source, _, template) = source();
+    let (_target, options) = destination(&template);
+    let held = acquire_runtime_store_owner_lock(&options.root).unwrap();
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    assert_fresh(&options);
+    drop(held);
+    initialize(&options.root, "wrong-runtime", WORKSPACE, NEW_URL);
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    initialize(&options.root, RUNTIME, WORKSPACE, "http://127.0.0.1:23456");
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    fs::remove_file(options.root.join("auth.toml")).unwrap();
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    initialize(&options.root, RUNTIME, WORKSPACE, NEW_URL);
+    fs::create_dir(options.root.join(WORKERS_DIR)).unwrap();
+    assert!(template.install(&options, request(), &resolved()).is_err());
+}
+
+#[test]
+fn extraction_rejects_nonstopped_pending_schema_and_metadata_mismatch() {
+    for change in [
+        "running",
+        "pending",
+        "guard-pending",
+        "attachment",
+        "schema",
+        "worker-id",
+        "metadata-workspace",
+        "metadata-session",
+        "artifact",
+        "secret",
+        "url",
+        "path",
+    ] {
+        let (_source, worker, _) = source();
+        let mut aggregate: WorkerAggregateRecord =
+            decode(&fs::read(worker.join(WORKER_FILE)).unwrap()).unwrap();
+        match change {
+            "running" => aggregate.status = WorkerStatus::Running,
+            "pending" => {
+                aggregate.execution_state = WorkerExecutionStateRecord::ReconciliationRequired(
+                    PersistedWorkerLifecycleOperation::Restore(PersistedWorkerRestoreOperation {
+                        operation_id: WorkerLifecycleOperationId::new(),
+                        mode: PersistedWorkerRestoreMode::Explicit,
+                        commit: crate::runtime::RestoreCommitFact::NotCommitted,
+                        source_repair_possible: false,
+                        request: request(),
+                        binding: None,
+                        last_settled_status: WorkerStatus::Stopped,
+                    }),
+                )
+            }
+            "guard-pending" => aggregate.restore_guard = serde_json::from_str(r#"{"observation_token":"old-token","active_request_id":"pending-restore","owners":{}}"#).unwrap(),
+            "attachment" => aggregate.logical_workdir_attachments.push(LogicalWorkdirAttachment {
+                alias: workdir::WorkdirAttachmentAlias::new("source").unwrap(),
+                working_directory_id: "old-workdir".into(),
+                capabilities: workdir::WorkdirSessionCapabilities::READ_ONLY,
+            }),
+            "schema" => aggregate.schema_version += 1,
+            "worker-id" => aggregate.worker_id = WorkerId::now_v7(),
+            "metadata-workspace" | "metadata-session" => {
+                let mut metadata: WorkerMetadata =
+                    decode(&fs::read(worker.join(WORKER_METADATA_FILE)).unwrap()).unwrap();
+                if change == "metadata-workspace" {
+                    metadata.workspace_id = Some("wrong-workspace".into());
+                } else {
+                    metadata.active.as_mut().unwrap().session_id = uuid::Uuid::now_v7();
+                }
+                atomic_write_json(
+                    &worker.join(WORKER_METADATA_FILE),
+                    &metadata,
+                    "test mismatch",
+                )
+                .unwrap();
+            }
+            "artifact" => {
+                fs::create_dir_all(worker.join("session/artifacts/paste")).unwrap();
+                fs::write(
+                    worker.join("session/artifacts/paste/payload"),
+                    "unsupported",
+                )
+                .unwrap();
+            }
+            "secret" | "url" | "path" => {
+                let segment = fs::read_dir(worker.join("session/segments"))
+                    .unwrap()
+                    .next()
+                    .unwrap()
+                    .unwrap()
+                    .path();
+                let replacement = match change {
+                    "secret" => "Bearer secret",
+                    "url" => "https://old-host.invalid",
+                    _ => "/old/host/path",
+                };
+                let text = fs::read_to_string(&segment)
+                    .unwrap()
+                    .replace("artificial fixture reply", replacement);
+                fs::write(segment, text).unwrap();
+            }
+            _ => unreachable!(),
+        }
+        atomic_write_json(&worker.join(WORKER_FILE), &aggregate, "test aggregate").unwrap();
+        assert!(
+            StoppedWorkerTemplate::extract(&worker).is_err(),
+            "accepted {change}"
+        );
+    }
+    let (_source, worker, _) = source();
+    let held =
+        acquire_runtime_store_owner_lock(worker.parent().unwrap().parent().unwrap()).unwrap();
+    assert!(StoppedWorkerTemplate::extract(&worker).is_err());
+    drop(held);
+}
+
+#[test]
+fn schema_v12_settled_template_installs_current_aggregate_without_rewriting_session() {
+    let (_source, _, mut template) = source();
+    template.runtime_schema_version = RESTORE_ADMISSION_SCHEMA_VERSION;
+    let bytes = serde_json::to_vec(&template).unwrap();
+    let (_target, options) = destination(&template);
+    template.install(&options, request(), &resolved()).unwrap();
+    assert_eq!(serde_json::to_vec(&template).unwrap(), bytes);
+    let worker = options.root.join(WORKERS_DIR).join(worker_id().to_string());
+    assert_eq!(
+        fs::read(worker.join("session/session.json")).unwrap(),
+        template.session_manifest
+    );
+    assert_eq!(
+        fs::read(worker.join(format!("session/segments/{}.jsonl", template.segment_id))).unwrap(),
+        template.segment_jsonl
+    );
+    let aggregate: WorkerAggregateRecord =
+        read_bounded_json(&worker.join(WORKER_FILE), "read installed fixture").unwrap();
+    let saved = aggregate.validate(&worker.join(WORKER_FILE)).unwrap();
+    assert_eq!(saved.status, WorkerStatus::Stopped);
+    assert!(matches!(
+        saved.execution_state,
+        PersistedWorkerExecutionState::Available(_)
+    ));
+    // Unsupported template schemas do not enter the versioned reader.
+    template.runtime_schema_version = RESTORE_ADMISSION_SCHEMA_VERSION - 1;
+    assert!(template.validate_shape().is_err());
+    template.runtime_schema_version = SCHEMA_VERSION + 1;
+    assert!(template.validate_shape().is_err());
+}
+
+#[test]
+fn fresh_manifest_and_memory_settings_are_required() {
+    let (_source, _, template) = source();
+    let (_target, options) = destination(&template);
+    let mut manifest = resolved();
+    manifest.worker.name = "wrong-worker".into();
+    assert!(template.install(&options, request(), &manifest).is_err());
+    let mut binding = request();
+    binding.memory_settings = None;
+    assert!(template.install(&options, binding, &resolved()).is_err());
+    let mut binding = request();
+    binding.memory_settings.as_mut().unwrap().workspace_id = "other-workspace".into();
+    assert!(template.install(&options, binding, &resolved()).is_err());
+    assert_fresh(&options);
+}
+
+#[test]
+fn settled_checkpoints_preserve_receipts_but_refuse_pending_work() {
+    let mut checkpoint = settled_checkpoint();
+    assert!(validate_settled_activation(&checkpoint_extension(&checkpoint)).is_ok());
+    checkpoint.pending.push(1);
+    assert!(validate_settled_activation(&checkpoint_extension(&checkpoint)).is_err());
+    checkpoint.pending.clear();
+    checkpoint.activating = Some(1);
+    assert!(validate_settled_activation(&checkpoint_extension(&checkpoint)).is_err());
+    checkpoint.activating = None;
+    checkpoint.invocation_receipts.push(1);
+    assert!(validate_settled_activation(&checkpoint_extension(&checkpoint)).is_err());
+    checkpoint.invocation_receipts.clear();
+    checkpoint.invocation_recovery_blocked = true;
+    assert!(validate_settled_activation(&checkpoint_extension(&checkpoint)).is_err());
+}
+
+// Deliberately corrupt only test-owned synthetic source files, never the lab.
+fn assert_rejected_session_bytes(worker: &Path, template: &StoppedWorkerTemplate, bytes: Vec<u8>) {
+    let path = worker.join(format!("session/segments/{}.jsonl", template.segment_id()));
+    fs::write(path, &bytes).unwrap();
+    let extraction_error = StoppedWorkerTemplate::extract(worker)
+        .err()
+        .expect("invalid source accepted");
+    assert!(
+        !extraction_error
+            .to_string()
+            .contains("synthetic-discarded-material")
+    );
+    let (_target, options) = destination(template);
+    let mut invalid = template.clone();
+    invalid.segment_jsonl = bytes;
+    let installation_error = invalid
+        .install(&options, request(), &resolved())
+        .unwrap_err();
+    assert!(
+        !installation_error
+            .to_string()
+            .contains("synthetic-discarded-material")
+    );
+    assert_fresh(&options);
+}
+
+fn synthetic_logs(template: &StoppedWorkerTemplate) -> Vec<LogEntry> {
+    template
+        .segment_jsonl
+        .split(|byte| *byte == b'\n')
+        .filter(|line| !line.is_empty())
+        .map(|line| serde_json::from_slice(line).unwrap())
+        .collect()
+}
+
+fn encode_synthetic_logs(logs: &[LogEntry]) -> Vec<u8> {
+    let mut bytes = Vec::new();
+    for log in logs {
+        serde_json::to_writer(&mut bytes, log).unwrap();
+        bytes.push(b'\n');
+    }
+    bytes
+}
+
+#[tokio::test]
+async fn missing_system_prompt_matches_worker_restore_rejection() {
+    let (_source, worker_dir, template) = source();
+    let mut logs = synthetic_logs(&template);
+    assert!(session_store::collect_state(&logs).system_prompt.is_some());
+    let LogEntry::AnnotatedSegmentStart { system_prompt, .. } = &mut logs[0] else {
+        panic!("missing seed")
+    };
+    *system_prompt = None;
+    assert_rejected_session_bytes(&worker_dir, &template, encode_synthetic_logs(&logs));
+    // The owning Worker constructor rejects this before model-client creation,
+    // allocation, or execution. Do not start a controller or submit any input.
+    let store = session_store::CombinedStore::new(
+        WorkerSessionStore::open_read_only(worker_dir.join("session")).unwrap(),
+        session_store::WorkerAggregateStore::open_read_only(&worker_dir, template.worker_name())
+            .unwrap(),
+    );
+    let restored = worker::Worker::restore_from_manifest(
+        template.session_id(),
+        template.segment_id(),
+        resolved(),
+        store,
+        worker::PromptCatalogSource::builtins_only(),
+    )
+    .await;
+    assert!(matches!(
+        restored,
+        Err(worker::WorkerError::SegmentSystemPromptMissing { .. })
+    ));
+}
+
+#[test]
+fn malformed_submission_and_notification_receipts_fail_extraction_and_install() {
+    let (_source, worker, template) = source();
+    for notification in [false, true] {
+        let mut checkpoint = settled_checkpoint();
+        if notification {
+            checkpoint.notification_receipts =
+                vec![ReceiptFixture::Malformed(1), ReceiptFixture::Malformed(2)];
+        } else {
+            checkpoint.receipts = vec![ReceiptFixture::Malformed(1), ReceiptFixture::Malformed(2)];
+        }
+        let mut logs = synthetic_logs(&template);
+        for log in &mut logs {
+            if let LogEntry::AnnotatedUserInput { extensions, .. } = log {
+                *extensions = vec![checkpoint_extension(&checkpoint)];
+            }
+        }
+        assert_rejected_session_bytes(&worker, &template, encode_synthetic_logs(&logs));
+    }
+}
+
+#[test]
+fn unknown_fields_at_every_retained_record_depth_are_rejected_without_secret_diagnostics() {
+    let (_source, worker, template) = source();
+    let original = String::from_utf8(template.segment_jsonl.clone()).unwrap();
+    // These values are harmless synthetic markers, not lab/private material.
+    // The compatibility decoder normally discards many of these fields.
+    let locations = [
+        r#""kind":"annotated_segment_start""#,
+        r#""kind":"invoke""#,
+        r#""kind":"annotated_user_input""#,
+        r#""kind":"annotated_assistant_item""#,
+        r#""kind":"llm_usage""#,
+        r#""kind":"turn_end""#,
+        r#""kind":"run_completed""#,
+        r#""max_tokens":null"#,
+        r#""kind":"message","role":"assistant""#,
+        r#""kind":"text","text":"artificial fixture reply""#,
+        r#""kind":"text","content":"Harmless artificial fixture input.""#,
+        r#""entry_id":"fixed-entry-id""#,
+        r#""kind":"model_output""#,
+        r#""worker_id":"01a1657d-b762-7791-87ae-a49f28cadbce""#,
+        r#""domain":"worker.pending_activations.v1""#,
+        r#""next_activation_sequence":2"#,
+        r#""submission_request_id":"completed-submission-request""#,
+        r#""notification_request_id":"completed-notification-request""#,
+    ];
+    for location in locations {
+        assert!(original.contains(location));
+        for extra in [
+            r#", "reasoning":{"text":"synthetic-discarded-material"}"#,
+            r#", "encrypted_content":"synthetic-discarded-material""#,
+        ] {
+            let altered = original.replacen(location, &format!("{location}{extra}"), 1);
+            assert_rejected_session_bytes(&worker, &template, altered.into_bytes());
+        }
+    }
+    let duplicate = original.replacen(
+        r#""role":"assistant""#,
+        r#""role":"assistant","role":"assistant""#,
+        1,
+    );
+    assert_rejected_session_bytes(&worker, &template, duplicate.into_bytes());
+}
+
+#[test]
+fn opaque_feature_payloads_cannot_bypass_strict_session_validation() {
+    let (_source, worker, template) = source();
+    let mut logs = synthetic_logs(&template);
+    for log in &mut logs {
+        if let LogEntry::AnnotatedUserInput { history, .. } = log {
+            if let session_store::LoggedItem::Message { role, .. } = &mut history[0].item {
+                *role = session_store::LoggedRole::System;
+            }
+            history[0].metadata.input_metadata = Some(
+                session_store::HistoryInputMetadata::new(
+                    session_store::HistoryInputUsePolicy {
+                        recording_material: session_store::InputEvidenceUse::Exclude,
+                        memory_evidence: session_store::InputEvidenceUse::Exclude,
+                        observation: session_store::InputObservationUse::Hidden,
+                        compaction: session_store::InputCompactionUse::Retain,
+                    },
+                    Some(
+                        session_store::FeatureInputPayload::encode(
+                            "fixture.test",
+                            1,
+                            &BTreeMap::from([("reasoning", "harmless synthetic payload")]),
+                        )
+                        .unwrap(),
+                    ),
+                    None,
+                )
+                .unwrap(),
+            );
+        }
+    }
+    assert_rejected_session_bytes(&worker, &template, encode_synthetic_logs(&logs));
+}
+
+#[cfg(unix)]
+#[test]
+fn refuses_source_and_destination_symlinks() {
+    let (_source, worker, template) = source();
+    let (_target, options) = destination(&template);
+    std::os::unix::fs::symlink(worker.join(WORKER_FILE), options.root.join("linked")).unwrap();
+    assert!(template.install(&options, request(), &resolved()).is_err());
+    fs::remove_file(options.root.join("linked")).unwrap();
+    std::os::unix::fs::symlink(worker.join(WORKER_FILE), worker.join("runs/link")).unwrap();
+    assert!(StoppedWorkerTemplate::extract(&worker).is_err());
+}
+
+/// Opt-in only: source path supplied explicitly; assertions never display content.
+#[test]
+#[ignore = "requires the approved artificial lab aggregate via YOI_STOPPED_FIXTURE_SOURCE"]
+fn approved_lab_extract_install() {
+    let path = PathBuf::from(
+        std::env::var_os("YOI_STOPPED_FIXTURE_SOURCE").expect("explicit source required"),
+    );
+    let template = StoppedWorkerTemplate::extract(&path).unwrap();
+    let (_target, options) = destination(&template);
+    let mut bind = request();
+    bind.worker_id = template.worker_id();
+    bind.workspace_api.as_mut().unwrap().workspace_id = template.workspace_id().into();
+    bind.memory_settings.as_mut().unwrap().workspace_id = template.workspace_id().into();
+    let mut manifest = resolved();
+    manifest.worker.name = template.worker_name();
+    template.install(&options, bind, &manifest).unwrap();
+    let installed = options
+        .root
+        .join(WORKERS_DIR)
+        .join(template.worker_id().to_string());
+    for relative in [
+        "session/session.json".to_string(),
+        format!("session/segments/{}.jsonl", template.segment_id()),
+    ] {
+        assert!(
+            fs::read(path.join(&relative)).unwrap() == fs::read(installed.join(relative)).unwrap()
+        );
+    }
+    let runtime = crate::Runtime::with_fs_store(options).unwrap();
+    assert_eq!(runtime.list_workers().unwrap().len(), 1);
+    assert_eq!(runtime.summary().unwrap().active_worker_count, 0);
+}

crates/worker-runtime/src/worker_backend/deferred_workdir.rs

@@ -193,6 +193,21 @@ mod tests {
     use super::*;
     use std::sync::atomic::{AtomicUsize, Ordering};
 
+    fn short_tempdir() -> tempfile::TempDir {
+        let target = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
+            .parent()
+            .unwrap()
+            .parent()
+            .unwrap()
+            .join("target");
+        std::fs::create_dir_all(&target).unwrap();
+        tempfile::Builder::new()
+            .prefix("d")
+            .rand_bytes(3)
+            .tempdir_in(target)
+            .unwrap()
+    }
+
     fn read_request() -> ReadRequest {
         ReadRequest {
             path: WorkdirPath::new("README.md").unwrap(),
@@ -225,9 +240,60 @@ mod tests {
         assert_eq!(attempts.load(Ordering::SeqCst), 0);
     }
 
+    #[tokio::test]
+    async fn router_cleanup_of_unopened_session_never_connects_or_follows_reused_alias() {
+        let attempts = Arc::new(AtomicUsize::new(0));
+        let observed = attempts.clone();
+        let session = DeferredWorkdirSession::handle(
+            "old",
+            WorkdirSessionCapabilities::READ_ONLY,
+            move || {
+                observed.fetch_add(1, Ordering::SeqCst);
+                Err("offline".into())
+            },
+        );
+        let router = WorkdirSessionRouter::new();
+        let alias = WorkdirAttachmentAlias::new("w").unwrap();
+        let routed = router.attach(alias.clone(), session.clone()).unwrap();
+        assert!(routed.subscribe_command_events().is_some());
+        assert!(routed.command_snapshot().is_empty());
+        router.close_all().await.unwrap();
+        router.close_all().await.unwrap();
+        assert!(router.aliases().is_empty());
+        assert!(matches!(
+            session.read(read_request()).await,
+            Err(WorkdirError::SessionClosed)
+        ));
+        assert_eq!(attempts.load(Ordering::SeqCst), 0);
+
+        let new_attempts = Arc::new(AtomicUsize::new(0));
+        let new_observed = new_attempts.clone();
+        let replacement = DeferredWorkdirSession::handle(
+            "new",
+            WorkdirSessionCapabilities::READ_ONLY,
+            move || {
+                new_observed.fetch_add(1, Ordering::SeqCst);
+                Err("offline".into())
+            },
+        );
+        let new_routed = router.attach(alias, replacement).unwrap();
+        // A retained old connection cannot open the replacement provider.
+        assert!(routed.read(read_request()).await.is_err());
+        assert_eq!(new_attempts.load(Ordering::SeqCst), 0);
+        assert!(matches!(
+            new_routed.read(read_request()).await,
+            Err(ref error) if matches!(error.classification_source(), WorkdirError::Unavailable(_))
+        ));
+        assert_eq!(new_attempts.load(Ordering::SeqCst), 1);
+        // Even after a failed open, cleanup must not try connecting again.
+        router.close_all().await.unwrap();
+        assert_eq!(new_attempts.load(Ordering::SeqCst), 1);
+        assert_eq!(attempts.load(Ordering::SeqCst), 0);
+    }
+
     #[tokio::test]
     async fn failed_open_retries_on_use_and_keeps_read_only_capabilities() {
-        let temp = tempfile::tempdir().unwrap();
+        let temp = short_tempdir();
         std::fs::write(temp.path().join("README.md"), "restored").unwrap();
         let root = temp.path().to_path_buf();
         let attempts = Arc::new(AtomicUsize::new(0));

crates/worker-runtime/src/worker_backend/restore_recovery.rs

@@ -0,0 +1,185 @@
+//! Explicit, operation-owned repair of saved Session material. This is not
+//! read-only candidate preparation: the canonical Session/attribution is the
+//! receipt for replay after a lost response, and effects are reported as such.
+
+use super::*;
+
+impl ProfileRuntimeWorkerFactory {
+    pub(super) async fn recover_saved_restore_source(
+        &self,
+        request: &WorkerExecutionRestoreRequest,
+    ) -> Result<bool, RuntimeRestoreFailure> {
+        let cause = |cause| RuntimeRestoreFailure::started(RuntimeRestoreStage::Recovery, cause);
+        let name = Self::runtime_worker_name_for_ref(&request.worker_ref);
+        let root = self
+            .worker_aggregate_dir(&request.worker_ref)
+            .map_err(|_| cause("restore recovery aggregate unavailable"))?;
+        let metadata_store = WorkerAggregateStore::open_read_only(&root, &name)
+            .map_err(|_| cause("restore recovery metadata open failed"))?;
+        let metadata = metadata_store
+            .read_by_name(&name)
+            .map_err(|_| cause("restore recovery metadata read failed"))?
+            .ok_or_else(|| cause("restore recovery metadata missing"))?;
+        let active = metadata
+            .active
+            .as_ref()
+            .ok_or_else(|| cause("restore recovery Session missing"))?;
+        let pending = active.segment_id.is_none();
+        let incomplete = request.request.subjektiv_attached
+            && !matches!(
+                metadata.subjektiv_session_attribution,
+                Some(session_store::SubjektivSessionAttributionState::Confirmed { .. })
+            );
+        if !pending && !incomplete {
+            return Ok(false);
+        }
+        request
+            .context
+            .authorize_restore_source_repair(request.operation_id)
+            .map_err(|_| cause("canonical source repair authority is not durably confirmed"))?;
+        // The accepted Runtime operation already owns this Session. Its stable
+        // Session identity and sole committed seed, not a new restore request,
+        // are used to reconcile an interrupted first-head publication.
+        let session_store = WorkerSessionStore::new_with_retention(
+            root.join("session"),
+            self.session_retention()
+                .map_err(|_| cause("restore recovery retention unavailable"))?,
+        )
+        .map_err(|_| cause("restore recovery Session open failed"))?;
+        let (manifest, mut loader) = Self::manifest_for_restore(&metadata, &request.request)
+            .map_err(|_| cause("restore recovery saved manifest invalid"))?;
+        let workspace = RuntimeWorkspaceBackendRef::from_worker_request(
+            &request.request,
+            self.runtime_id.as_deref(),
+        )
+        .map_err(cause)?;
+        #[cfg(feature = "http-server")]
+        let request_client = request
+            .request
+            .workspace_api
+            .as_ref()
+            .and_then(|api| self.workspace_request_clients.get(&api.workspace_id));
+        let context = workspace.worker_context(
+            #[cfg(feature = "http-server")]
+            &request.worker_ref,
+            #[cfg(feature = "http-server")]
+            request.workspace_scope.as_ref(),
+            #[cfg(feature = "http-server")]
+            self.worker_mutation_identity.as_ref(),
+            #[cfg(feature = "http-server")]
+            request_client,
+            #[cfg(feature = "http-server")]
+            self.embedded_worker_mutation_dispatcher.as_ref(),
+            #[cfg(feature = "http-server")]
+            Some(self.prompt_projection_cache.clone()),
+        );
+        if pending && context.workspace_id().is_some() {
+            let bundle = request
+                .config_bundle
+                .as_ref()
+                .ok_or_else(|| cause("pending recovery launch material missing"))?;
+            let projection = self
+                .observe_bundle_prompt_projection(
+                    bundle,
+                    request
+                        .request
+                        .workspace_api
+                        .as_ref()
+                        .map(|api| api.workspace_id.as_str()),
+                )
+                .map_err(|_| cause("pending recovery Prompt projection invalid"))?
+                .ok_or_else(|| cause("pending recovery Prompt projection missing"))?;
+            loader = loader.with_effective_catalog(projection.projection.catalog.clone());
+        }
+        self.failed_restore_resources
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .insert(
+                request.worker_ref.clone(),
+                FailedRestoreResources {
+                    operation_id: request.operation_id,
+                    sessions: Vec::new(),
+                    candidate_workdir_cleanup: None,
+                    allocation: None,
+                    feature_tasks: None,
+                },
+            );
+        let store = CombinedStore::new(session_store, metadata_store);
+        let worker_result = Worker::restore_pending_from_worker_metadata_with_context(
+            &name,
+            manifest,
+            store,
+            loader,
+            context,
+            WorkerFilesystemAuthority::None,
+        )
+        .await;
+        let mut worker = worker_result.map_err(|error| {
+            self.retain_construction_failure(request, &error);
+            cause(restore_worker_cause(&error))
+        })?;
+        let tasks = worker.retain_startup_feature_cleanup();
+        let sessions = worker.workdir_sessions();
+        let allocation = worker.restore_allocation_cleanup_handle();
+        self.failed_restore_resources
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .insert(
+                request.worker_ref.clone(),
+                FailedRestoreResources {
+                    operation_id: request.operation_id,
+                    sessions: vec![sessions.clone()],
+                    candidate_workdir_cleanup: None,
+                    allocation: allocation.clone(),
+                    feature_tasks: Some(tasks.clone()),
+                },
+            );
+        let result = async {
+            if let Some(binding) = request.request.backend_job.as_ref() {
+                worker
+                    .bind_backend_job(binding.clone())
+                    .map_err(|_| cause("recovery Backend Job binding failed"))?;
+            }
+            worker
+                .finalize_subjektiv_session_attribution(
+                    request.request.subjektiv_attached,
+                    SubjektivSessionAttributionLifecycle::RestoredSession,
+                )
+                .map_err(|error| cause(restore_worker_cause(&error)))?;
+            if pending {
+                worker::WorkerController::recover_pending_head(&mut worker, &root)
+                    .await
+                    .map_err(|_| cause("pending Session head recovery failed"))?;
+            }
+            Ok::<_, RuntimeRestoreFailure>(())
+        }
+        .await;
+        // Never release allocation while the Worker guard is attached. Retained
+        // handles in the same operation remain available on every cleanup error.
+        drop(worker);
+        let cleanup = async {
+            tasks
+                .shutdown()
+                .await
+                .map_err(|_| cause("recovery Feature settlement failed"))?;
+            sessions
+                .close_all()
+                .await
+                .map_err(|_| cause("recovery Workdir close failed"))?;
+            if let Some(allocation) = allocation {
+                allocation
+                    .release()
+                    .map_err(|_| cause("recovery allocation release failed"))?;
+            }
+            Ok::<_, RuntimeRestoreFailure>(())
+        }
+        .await;
+        result?;
+        cleanup?;
+        self.failed_restore_resources
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .remove(&request.worker_ref);
+        Ok(true)
+    }
+}

crates/worker-runtime/src/worker_backend_restore_registration_tests.rs

@@ -0,0 +1,927 @@
+// Included by the production Profile publication suite. Each fault is at a
+// distinct partial-publication boundary; no factory result is substituted.
+
+fn publication_gate<T>() -> (
+    mpsc::Sender<()>,
+    mpsc::Receiver<T>,
+    crate::execution::RestoreTestGate<T>,
+) {
+    let (release, receiver) = mpsc::channel();
+    let (entered, evidence) = mpsc::channel();
+    (
+        release,
+        evidence,
+        crate::execution::RestoreTestGate {
+            entered,
+            release: receiver,
+        },
+    )
+}
+
+fn assert_single_publication(fixture: &Fixture, expected: usize) {
+    let faults = &fixture.backend.factory.restore_faults;
+    assert_eq!(
+        faults.executions.load(Ordering::SeqCst),
+        expected,
+        "construction reran"
+    );
+    assert_eq!(
+        faults.effects_calls.load(Ordering::SeqCst),
+        expected,
+        "restore effects reran"
+    );
+    assert_eq!(
+        faults.publication_attempts.load(Ordering::SeqCst),
+        expected,
+        "Controller start reran"
+    );
+    fixture.assert_quiet();
+}
+
+fn assert_real_attach(
+    fixture: &Fixture,
+    handle: &WorkerHandle,
+) -> crate::execution::WorkerProtocolTransport {
+    let worker_ref = &fixture.request.worker_ref;
+    assert!(matches!(
+        fixture
+            .backend
+            .worker_session(crate::execution::WorkerSessionObservationRequest {
+                worker_ref: worker_ref.clone(),
+            }),
+        runtime_api::WorkerSessionAvailability::LiveProtocol
+    ));
+    let transport = fixture
+        .backend
+        .clone()
+        .attach_worker_protocol(worker_ref)
+        .unwrap();
+    transport.validate().unwrap();
+    let paired = subscribe_worker_protocol_session(handle);
+    assert_eq!(
+        serde_json::to_value(&transport.snapshot).unwrap(),
+        serde_json::to_value(paired.snapshot_event).unwrap()
+    );
+    assert!(
+        matches!(&transport.snapshot, Event::Snapshot { session, greeting, state, in_flight, .. }
+        if session == &session_store::public_snapshot::project_current_session_snapshot(&fixture.entries())
+        && !greeting.tools.is_empty() && state.state == protocol::WorkerState::Idle && in_flight.is_empty())
+    );
+    transport
+}
+
+fn observe_publication_markers(
+    fixture: &Fixture,
+    observations: tokio::sync::broadcast::Receiver<crate::observation::WorkerObservationEvent>,
+    markers: &'static [&'static str],
+) -> tokio::sync::broadcast::Receiver<crate::observation::WorkerObservationEvent> {
+    fixture
+        .backend
+        .run_on_adapter_runtime(async move {
+            let mut observations = observations;
+            let mut counts = vec![0; markers.len()];
+            let mut snapshots = 0;
+            tokio::time::timeout(WAIT, async {
+                loop {
+                    let observed = observations.recv().await.unwrap();
+                    if matches!(observed.payload, Event::Snapshot { .. }) {
+                        snapshots += 1;
+                    }
+                    for (index, marker) in markers.iter().enumerate() {
+                        if is_marker(&observed.payload, marker) {
+                            counts[index] += 1;
+                        }
+                    }
+                    if counts.last() == Some(&1) {
+                        break;
+                    }
+                }
+            })
+            .await
+            .map_err(|_| format!("lost original subscription markers: {counts:?}"))?;
+            assert_eq!(
+                counts,
+                vec![1; markers.len()],
+                "lost or duplicated prefix on original receiver"
+            );
+            assert_eq!(snapshots, 1, "initial full snapshot was lost or replayed");
+            while let Ok(event) = observations.try_recv() {
+                assert!(
+                    !markers
+                        .iter()
+                        .any(|marker| is_marker(&event.payload, marker)),
+                    "duplicate after tail fence"
+                );
+            }
+            Ok(observations)
+        })
+        .unwrap()
+}
+
+fn partial_registration_recovers_same_controller(hub: bool) {
+    let mut fixture = Fixture::no_model();
+    Arc::get_mut(&mut fixture.backend)
+        .unwrap()
+        .spawn_restore_timeout = Duration::from_millis(250);
+    let runtime = stopped_runtime(&fixture);
+    let pause = runtime.test_pause_restore_continuation();
+    let worker_ref = fixture.request.worker_ref.clone();
+    let scope = fixture.request.workspace_scope.as_ref().unwrap();
+    let request = runtime.test_restore_request(&worker_ref);
+    let observations = runtime.subscribe_worker_observation().unwrap();
+    let (started_release, started_entered) = activation_gate(&fixture, true);
+    let first = runtime
+        .restore_worker_operation_scoped(scope, &worker_ref, request.clone())
+        .unwrap();
+    assert_eq!(
+        first.state,
+        server_api::WorkerRestoreState::ReconciliationRequired
+    );
+    let started = started_entered.recv_timeout(WAIT).unwrap();
+    let pending = assert_activation_owned(&fixture, &runtime, &request.request_id, &started);
+    let old = started.handle.as_ref().unwrap().clone();
+    let canonical = fixture.canonical_files();
+    alert_marker(&old, "registration-original-paired-prefix");
+    let (release, entered, fault) = publication_gate();
+    if hub {
+        *fixture
+            .backend
+            .factory
+            .observation_hub
+            .registration_gate
+            .lock()
+            .unwrap() = Some(RestoreHubRegistrationGate {
+            gate: fault,
+            fail: true,
+        });
+    } else {
+        *fixture.backend.bridge_fault.lock().unwrap() = Some(fault);
+    }
+    started_release.release();
+    drop(pause);
+    let boundary_handle = entered.recv_timeout(WAIT).unwrap();
+    assert!(boundary_handle.same_controller(&old));
+    let durable = durable_operation(&runtime, &worker_ref).unwrap();
+    assert_eq!(durable.operation_id, started.operation_id);
+    assert_eq!(durable.commit, crate::runtime::RestoreCommitFact::Committed);
+    assert!(restore_owner(&runtime, &worker_ref, &request.request_id)["receipt"].is_null());
+    if hub {
+        let (_, shared, _) = fixture
+            .backend
+            .factory
+            .observation_hub
+            .get(&worker_ref)
+            .unwrap();
+        assert!(
+            Arc::ptr_eq(&shared, &old.shared_state),
+            "real hub insert preceded fault"
+        );
+    } else {
+        assert!(
+            fixture
+                .backend
+                .factory
+                .observation_hub
+                .get(&worker_ref)
+                .is_none()
+        );
+    }
+    let (operation, context, uncertain) = runtime.test_pending_restore_context(&worker_ref);
+    assert_eq!(operation, started.operation_id);
+    assert!(!uncertain);
+    assert!(
+        context
+            .test_buffered_events()
+            .iter()
+            .any(|event| matches!(event, Event::Snapshot { .. }))
+    );
+    alert_marker(&old, "registration-during-fault");
+    let pause = runtime.test_pause_restore_continuation();
+    release.send(()).unwrap();
+    // This lock also observes the end of the actual registration attempt, not
+    // merely the test's entered notification. Its retained scope owns cleanup.
+    let execution = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(old.same_controller(&execution.handle));
+    assert_eq!(execution.restore_operation_id, Some(operation));
+    assert_eq!(
+        execution.tasks.tasks.lock().unwrap().len(),
+        if hub { 2 } else { 1 }
+    );
+    let bridge = execution.tasks.protocol_bridge.lock().unwrap();
+    if hub {
+        assert!(bridge.is_none());
+    } else {
+        let bridge = bridge.as_ref().unwrap();
+        assert!(
+            !bridge.snapshot_pending,
+            "snapshot already registered, do not replay it"
+        );
+        let streams = bridge.streams.lock().unwrap();
+        assert!(
+            streams.is_some(),
+            "original paired streams survived failed task admission"
+        );
+        assert_eq!(
+            serde_json::to_value(&streams.as_ref().unwrap().snapshot_event).unwrap(),
+            serde_json::to_value(&started.snapshot).unwrap()
+        );
+    }
+    drop(bridge);
+    assert!(
+        !old.protocol_is_closed(),
+        "publication failure must not clean a committed Controller"
+    );
+    assert_single_publication(&fixture, 1);
+    drop(pause);
+    runtime.test_wait_restore_continuation(&worker_ref, operation, WAIT);
+    assert!(durable_operation(&runtime, &worker_ref).is_none());
+    let owner = restore_owner(&runtime, &worker_ref, &request.request_id);
+    assert_eq!(
+        owner["operation_id"],
+        serde_json::to_value(operation).unwrap()
+    );
+    assert_eq!(owner["receipt"]["state"], "accepted");
+    assert!(
+        fixture
+            .backend
+            .pending_restore(&worker_ref)
+            .unwrap()
+            .is_none()
+    );
+    assert!(!pending.task_in_flight.load(Ordering::Acquire));
+    assert!(pending.activated.lock().unwrap().is_none());
+    assert!(pending.result.lock().unwrap().is_none());
+    assert!(pending.task.try_lock().unwrap().is_none());
+    assert_eq!(fixture.backend.workers.lock().unwrap().len(), 1);
+    assert!(old.same_controller(&fixture.backend.workers.lock().unwrap()[&worker_ref].handle));
+    let transport = assert_real_attach(&fixture, &old);
+    assert_eq!(fixture.canonical_files(), canonical);
+    assert_single_publication(&fixture, 1);
+    alert_marker(&old, "registration-tail-fence");
+    let observations = observe_publication_markers(
+        &fixture,
+        observations,
+        &[
+            "registration-original-paired-prefix",
+            "registration-during-fault",
+            "registration-tail-fence",
+        ],
+    );
+    assert_eq!(
+        runtime
+            .restore_worker_operation_scoped(scope, &worker_ref, request.clone())
+            .unwrap()
+            .state,
+        server_api::WorkerRestoreState::Accepted
+    );
+    assert_single_publication(&fixture, 1);
+    // Retain the original transport/handle and scope through Stop/replacement.
+    runtime
+        .stop_worker_scoped(scope, &worker_ref, None)
+        .unwrap();
+    assert!(old.protocol_is_closed());
+    assert!(transport.validate().is_err());
+    let replacement_request = runtime.test_restore_request(&worker_ref);
+    assert_eq!(
+        runtime
+            .restore_worker_operation_scoped(scope, &worker_ref, replacement_request.clone())
+            .unwrap()
+            .state,
+        server_api::WorkerRestoreState::Accepted
+    );
+    let replacement = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(!old.same_controller(&replacement.handle));
+    assert_ne!(replacement.restore_operation_id, Some(operation));
+    assert!(
+        transport
+            .dispatch(Method::Pause {
+                command: protocol::WorkerCommandEnvelope::new(801)
+            })
+            .is_err()
+    );
+    assert!(
+        transport
+            .dispatch(Method::Shutdown {
+                command: protocol::WorkerCommandEnvelope::new(802)
+            })
+            .is_err()
+    );
+    for method in [
+        Method::Pause {
+            command: protocol::WorkerCommandEnvelope::new(803),
+        },
+        Method::Shutdown {
+            command: protocol::WorkerCommandEnvelope::new(804),
+        },
+    ] {
+        let old_command = old.clone();
+        assert!(
+            fixture
+                .backend
+                .run_on_adapter_runtime(async move {
+                    old_command
+                        .send(method)
+                        .await
+                        .map_err(|error| error.to_string())
+                })
+                .is_err(),
+            "closed original Controller command must not route to replacement"
+        );
+    }
+    fixture
+        .backend
+        .factory
+        .deactivate_stopped_controller(&worker_ref, &old)
+        .unwrap();
+    let (_, shared, _) = fixture
+        .backend
+        .factory
+        .observation_hub
+        .get(&worker_ref)
+        .unwrap();
+    assert!(Arc::ptr_eq(&shared, &replacement.handle.shared_state));
+    alert_marker(&old, "registration-stopped-event-must-not-leak");
+    alert_marker(&replacement.handle, "registration-replacement-fence");
+    fixture
+        .backend
+        .run_on_adapter_runtime(async move {
+            let mut observations = observations;
+            tokio::time::timeout(WAIT, async {
+                loop {
+                    let event = observations.recv().await.unwrap();
+                    assert!(!is_marker(
+                        &event.payload,
+                        "registration-stopped-event-must-not-leak"
+                    ));
+                    if is_marker(&event.payload, "registration-replacement-fence") {
+                        break;
+                    }
+                }
+            })
+            .await
+            .map_err(|_| "replacement observation fence missing".to_string())?;
+            drop(transport);
+            Ok(())
+        })
+        .unwrap();
+    assert_single_publication(&fixture, 2);
+    assert_eq!(
+        runtime
+            .restore_worker_operation_scoped(scope, &worker_ref, request.clone())
+            .unwrap()
+            .state,
+        server_api::WorkerRestoreState::Accepted
+    );
+    assert_eq!(restore_owner_count(&runtime, &worker_ref), 2);
+    let current = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(current.handle.same_controller(&replacement.handle));
+    assert_eq!(
+        current.restore_operation_id,
+        replacement.restore_operation_id
+    );
+    assert_eq!(
+        replacement.handle.shared_state.snapshot().state,
+        protocol::WorkerState::Idle
+    );
+    drop(current);
+    drop(execution);
+    drop(replacement);
+    drop(pending);
+    runtime
+        .stop_worker_scoped(scope, &worker_ref, None)
+        .unwrap();
+    fixture.assert_released();
+}
+
+fn restore_owner_count(runtime: &EmbeddedRuntime, worker_ref: &WorkerRef) -> usize {
+    let saved = runtime
+        .test_runtime_store()
+        .read_worker_record(worker_ref)
+        .unwrap();
+    serde_json::to_value(saved.restore_guard).unwrap()["owners"]
+        .as_object()
+        .unwrap()
+        .len()
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn bridge_partial_snapshot_registration_retains_original_streams_and_autonomously_recovers() {
+    partial_registration_recovers_same_controller(false);
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn hub_partial_insert_registration_retains_original_controller_and_autonomously_recovers() {
+    partial_registration_recovers_same_controller(true);
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn context_partial_flush_retains_only_undelivered_queue_and_recovers_without_repost() {
+    let mut fixture = Fixture::no_model();
+    Arc::get_mut(&mut fixture.backend)
+        .unwrap()
+        .spawn_restore_timeout = Duration::from_millis(250);
+    let runtime = stopped_runtime(&fixture);
+    let pause = runtime.test_pause_restore_continuation();
+    let worker_ref = fixture.request.worker_ref.clone();
+    let scope = fixture.request.workspace_scope.as_ref().unwrap();
+    let request = runtime.test_restore_request(&worker_ref);
+    let observations = runtime.subscribe_worker_observation().unwrap();
+    let (started_release, started_entered) = activation_gate(&fixture, true);
+    assert_eq!(
+        runtime
+            .restore_worker_operation_scoped(scope, &worker_ref, request.clone())
+            .unwrap()
+            .state,
+        server_api::WorkerRestoreState::ReconciliationRequired
+    );
+    let started = started_entered.recv_timeout(WAIT).unwrap();
+    let pending = assert_activation_owned(&fixture, &runtime, &request.request_id, &started);
+    let old = started.handle.unwrap();
+    let (operation, context, _) = runtime.test_pending_restore_context(&worker_ref);
+    // Stop after the actual bridge and hub registration, without failing that
+    // boundary, so the flush begins with its real original full snapshot.
+    let (hub_release, hub_entered, hub_gate) = publication_gate();
+    *fixture
+        .backend
+        .factory
+        .observation_hub
+        .registration_gate
+        .lock()
+        .unwrap() = Some(RestoreHubRegistrationGate {
+        gate: hub_gate,
+        fail: false,
+    });
+    started_release.release();
+    drop(pause);
+    assert!(
+        hub_entered
+            .recv_timeout(WAIT)
+            .unwrap()
+            .same_controller(&old)
+    );
+    let buffered_prefix = context.test_buffered_events();
+    assert!(matches!(
+        buffered_prefix.first(),
+        Some(Event::Snapshot { .. })
+    ));
+    let prefix_length = buffered_prefix.len();
+    let marker = |text: &str| Event::Error {
+        code: protocol::ErrorCode::Internal,
+        message: text.into(),
+    };
+    context
+        .publish_protocol_event(marker("flush-delivered"))
+        .unwrap();
+    context
+        .publish_protocol_event(marker("flush-failed"))
+        .unwrap();
+    context
+        .publish_protocol_event(marker("flush-pending"))
+        .unwrap();
+    let (release, entered, gate) = publication_gate();
+    let (restored, restored_queue) = mpsc::channel();
+    context.test_fail_flush(crate::execution::RestoreFlushFault {
+        remaining_successes: prefix_length + 1,
+        gate,
+        restored,
+    });
+    drop(hub_release);
+    assert!(
+        matches!(entered.recv_timeout(WAIT).unwrap(), Event::Error { message, .. } if message == "flush-failed")
+    );
+    let partial_execution = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(old.same_controller(&partial_execution.handle));
+    assert_eq!(partial_execution.restore_operation_id, Some(operation));
+    assert_eq!(partial_execution.tasks.tasks.lock().unwrap().len(), 2);
+    let (_, shared, _) = fixture
+        .backend
+        .factory
+        .observation_hub
+        .get(&worker_ref)
+        .unwrap();
+    assert!(Arc::ptr_eq(&shared, &old.shared_state));
+    let pause = runtime.test_pause_restore_continuation();
+    context
+        .publish_protocol_event(marker("flush-arrived-during-error"))
+        .unwrap();
+    release.send(()).unwrap();
+    let restored = restored_queue.recv_timeout(WAIT).unwrap();
+    let undelivered: Vec<_> = restored
+        .iter()
+        .filter_map(|event| match event {
+            Event::Error { message, .. } if message.starts_with("flush-") => Some(message.as_str()),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(
+        undelivered,
+        [
+            "flush-failed",
+            "flush-pending",
+            "flush-arrived-during-error"
+        ]
+    );
+    assert!(
+        !restored
+            .iter()
+            .any(|event| matches!(event, Event::Snapshot { .. })),
+        "delivered snapshot must not rejoin the failed suffix"
+    );
+    assert_eq!(
+        serde_json::to_value(context.test_buffered_events()).unwrap(),
+        serde_json::to_value(&restored).unwrap()
+    );
+    assert_eq!(
+        durable_operation(&runtime, &worker_ref)
+            .unwrap()
+            .operation_id,
+        operation
+    );
+    assert!(restore_owner(&runtime, &worker_ref, &request.request_id)["receipt"].is_null());
+    // The next terminal gate observes successful retry only after the same
+    // queued owner has repaired its failed callback and remaining suffix.
+    let (terminal_release, terminal_entered, gate) = publication_gate();
+    runtime.test_terminal_receipt_gate(gate);
+    drop(pause);
+    assert_eq!(terminal_entered.recv_timeout(WAIT).unwrap(), operation);
+    assert!(context.test_buffered_events().is_empty());
+    assert!(restore_owner(&runtime, &worker_ref, &request.request_id)["receipt"].is_null());
+    assert_eq!(
+        durable_operation(&runtime, &worker_ref).unwrap().commit,
+        crate::runtime::RestoreCommitFact::Committed
+    );
+    let execution = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(old.same_controller(&execution.handle));
+    assert_eq!(execution.restore_operation_id, Some(operation));
+    drop(terminal_release); // RAII disconnect, not a second request.
+    runtime.test_wait_restore_continuation(&worker_ref, operation, WAIT);
+    assert_eq!(
+        restore_owner(&runtime, &worker_ref, &request.request_id)["receipt"]["state"],
+        "accepted"
+    );
+    assert_real_attach(&fixture, &old);
+    fixture
+        .backend
+        .run_on_adapter_runtime(async move {
+            let mut observations = observations;
+            let mut messages = Vec::new();
+            let mut snapshots = 0;
+            while let Ok(event) = observations.try_recv() {
+                match event.payload {
+                    Event::Error { message, .. } if message.starts_with("flush-") => {
+                        messages.push(message)
+                    }
+                    Event::Snapshot { .. } => snapshots += 1,
+                    _ => {}
+                }
+            }
+            assert_eq!(
+                messages,
+                [
+                    "flush-delivered",
+                    "flush-failed",
+                    "flush-pending",
+                    "flush-arrived-during-error"
+                ]
+            );
+            assert_eq!(snapshots, 1);
+            Ok(())
+        })
+        .unwrap();
+    assert_single_publication(&fixture, 1);
+    drop(partial_execution);
+    drop(execution);
+    drop(pending);
+    runtime
+        .stop_worker_scoped(scope, &worker_ref, None)
+        .unwrap();
+    fixture.assert_released();
+}
+
+#[derive(Clone, Copy, Debug)]
+enum TerminalFault {
+    BeforeRename,
+    AfterRename,
+    SyncUnknown,
+    ReadUnknown,
+}
+
+fn published_terminal_receipt_fault(fault: TerminalFault) {
+    let mut fixture = Fixture::no_model();
+    // Reuse the existing ordinary non-network model boundary, not the product
+    // Codex profile. All fault/convergence assertions precede the sole input.
+    let script = MockClient::sequential(vec![MockResponse::Complete(simple_text_events())]);
+    let injected = script.clone();
+    let factory = fixture
+        .backend
+        .factory
+        .as_ref()
+        .clone()
+        .with_spawn_model_client_factory(move || Box::new(injected.clone()));
+    fixture.backend = Arc::new(WorkerRuntimeExecutionBackend::new(factory).unwrap());
+    let runtime = Arc::new(stopped_runtime(&fixture));
+    let pause = runtime.test_pause_restore_continuation();
+    let worker_ref = fixture.request.worker_ref.clone();
+    let scope = fixture.request.workspace_scope.clone().unwrap();
+    let request = runtime.test_restore_request(&worker_ref);
+    let observations = runtime.subscribe_worker_observation().unwrap();
+    let mut catalog = runtime
+        .subscribe_event_selector_scoped(
+            &scope,
+            protocol::subscription::EventSubscriptionSelector::RuntimeWorkers,
+        )
+        .unwrap();
+    let (release, entered, gate) = publication_gate();
+    runtime.test_terminal_receipt_gate(gate);
+    let restoring = {
+        let runtime = runtime.clone();
+        let worker_ref = worker_ref.clone();
+        let scope = scope.clone();
+        let request = request.clone();
+        std::thread::spawn(move || {
+            runtime.restore_worker_operation_scoped(&scope, &worker_ref, request)
+        })
+    };
+    let operation = entered.recv_timeout(WAIT).unwrap();
+    // Restore legitimately repairs derived metadata before publication. Receipt
+    // failures must not rerun those effects; compare from this actual boundary.
+    let canonical = fixture.canonical_files();
+    let execution = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    let old = execution.handle.clone();
+    assert_eq!(execution.restore_operation_id, Some(operation));
+    assert!(
+        fixture
+            .backend
+            .pending_restore(&worker_ref)
+            .unwrap()
+            .is_none()
+    );
+    assert_eq!(execution.tasks.tasks.lock().unwrap().len(), 2);
+    assert!(execution.tasks.protocol_bridge.lock().unwrap().is_none());
+    assert_eq!(
+        durable_operation(&runtime, &worker_ref)
+            .unwrap()
+            .operation_id,
+        operation
+    );
+    assert_eq!(
+        durable_operation(&runtime, &worker_ref).unwrap().commit,
+        crate::runtime::RestoreCommitFact::Committed
+    );
+    let owner = restore_owner(&runtime, &worker_ref, &request.request_id);
+    assert!(owner["receipt"].is_null());
+    let (same_operation, context, uncertain) = runtime.test_pending_restore_context(&worker_ref);
+    assert_eq!(same_operation, operation);
+    assert!(!uncertain);
+    assert!(context.test_buffered_events().is_empty());
+    let transport = assert_real_attach(&fixture, &old);
+    assert_single_publication(&fixture, 1);
+    alert_marker(&old, "terminal-before-receipt");
+    let store = runtime.test_runtime_store();
+    let reconcile_before = store.test_worker_reconcile_counts();
+    store.fail_next_worker_write(if matches!(fault, TerminalFault::BeforeRename) {
+        AtomicWriteFault::BeforeRename
+    } else {
+        AtomicWriteFault::AfterRename
+    });
+    match fault {
+        TerminalFault::SyncUnknown => store.fail_next_worker_reconcile_sync(),
+        TerminalFault::ReadUnknown => store.fail_next_worker_reconcile_read(),
+        _ => {}
+    }
+    release.send(()).unwrap();
+    let first = restoring.join().unwrap().unwrap();
+    let reconcile_after = store.test_worker_reconcile_counts();
+    let reconciliation = (
+        reconcile_after.0 - reconcile_before.0,
+        reconcile_after.1 - reconcile_before.1,
+    );
+    assert_eq!(
+        reconciliation,
+        match fault {
+            TerminalFault::BeforeRename => (0, 0),
+            TerminalFault::AfterRename => (1, 4), // canonical file plus Worker/workers/root directories
+            TerminalFault::SyncUnknown => (1, 1),
+            TerminalFault::ReadUnknown => (1, 0),
+        },
+        "{fault:?}: actual canonical read/sync attempts"
+    );
+    assert_eq!(
+        first.state,
+        if matches!(fault, TerminalFault::AfterRename) {
+            server_api::WorkerRestoreState::Accepted
+        } else {
+            server_api::WorkerRestoreState::ReconciliationRequired
+        },
+        "{fault:?}: {first:?}"
+    );
+    assert_eq!(fixture.canonical_files(), canonical);
+    assert!(!old.protocol_is_closed());
+    transport.validate().unwrap();
+    assert_single_publication(&fixture, 1);
+    let disk_owner = restore_owner(&runtime, &worker_ref, &request.request_id);
+    assert_eq!(disk_owner["operation_id"], owner["operation_id"]);
+    match fault {
+        TerminalFault::BeforeRename => {
+            assert!(disk_owner["receipt"].is_null());
+            assert!(!runtime.test_pending_restore_context(&worker_ref).2);
+        }
+        TerminalFault::AfterRename => {
+            assert_eq!(disk_owner["receipt"]["state"], "accepted");
+            assert!(durable_operation(&runtime, &worker_ref).is_none());
+        }
+        TerminalFault::SyncUnknown | TerminalFault::ReadUnknown => {
+            assert_eq!(
+                disk_owner["receipt"]["state"], "accepted",
+                "visible rename alone is not durable proof"
+            );
+            let (same_operation, _, fenced) = runtime.test_pending_restore_context(&worker_ref);
+            assert_eq!(same_operation, operation);
+            assert!(
+                fenced,
+                "unknown canonical outcome must retain the memory owner/write fence"
+            );
+            // Keep durability unproved for one competing Stop as well: it must
+            // fail before any Shutdown, never manufacture a Stopped judgment.
+            match fault {
+                TerminalFault::SyncUnknown => store.fail_next_worker_reconcile_sync(),
+                TerminalFault::ReadUnknown => store.fail_next_worker_reconcile_read(),
+                _ => unreachable!(),
+            }
+            assert!(matches!(
+                runtime.stop_worker_scoped(&scope, &worker_ref, None),
+                Err(crate::error::RuntimeError::StoreCommitOutcomeUnknown { .. })
+            ));
+            assert!(!old.protocol_is_closed());
+            assert!(!execution.shutdown_requested.load(Ordering::Acquire));
+            assert!(runtime.test_pending_restore_context(&worker_ref).2);
+            assert_single_publication(&fixture, 1);
+        }
+    }
+    // Recovery is autonomous after the failed write / confirmed unknown fence.
+    // No POST, fresh request, or input is needed to make it converge.
+    drop(pause);
+    if !matches!(fault, TerminalFault::AfterRename) {
+        runtime.test_wait_restore_continuation(&worker_ref, operation, WAIT);
+    }
+    assert!(durable_operation(&runtime, &worker_ref).is_none());
+    assert_eq!(
+        restore_owner(&runtime, &worker_ref, &request.request_id)["receipt"]["state"],
+        "accepted"
+    );
+    let current = fixture.backend.workers.lock().unwrap()[&worker_ref].clone();
+    assert!(old.same_controller(&current.handle));
+    assert_eq!(current.restore_operation_id, Some(operation));
+    assert_eq!(fixture.backend.workers.lock().unwrap().len(), 1);
+    assert_real_attach(&fixture, &old);
+    assert_single_publication(&fixture, 1);
+    // A canonical terminal reread may bypass the final redundant upsert. The
+    // original subscriber must still have the current state and observation token.
+    let fresh_catalog = runtime
+        .subscribe_event_selector_scoped(
+            &scope,
+            protocol::subscription::EventSubscriptionSelector::RuntimeWorkers,
+        )
+        .unwrap();
+    let protocol::subscription::SubscriptionSnapshot::Workers { workers } =
+        fresh_catalog.snapshot()
+    else {
+        panic!("Runtime catalog must return Workers");
+    };
+    assert_eq!(workers.len(), 1);
+    let mut last_applied = None;
+    use futures::FutureExt;
+    while let Some(update) = catalog.recv().now_or_never() {
+        match update.unwrap().payload {
+            protocol::subscription::SubscriptionEventPayload::WorkerUpserted { worker } => {
+                last_applied = Some(worker);
+            }
+            payload => panic!("unexpected catalog update: {payload:?}"),
+        }
+    }
+    assert!(
+        serde_json::to_value(last_applied.expect("restore must publish its catalog state"))
+            .unwrap()
+            == serde_json::to_value(&workers[0]).unwrap(),
+        "original catalog subscriber diverged after {fault:?}"
+    );
+    drop(fresh_catalog);
+    drop(catalog);
+    alert_marker(&old, "terminal-after-recovery");
+    let _observations = observe_publication_markers(
+        &fixture,
+        observations,
+        &["terminal-before-receipt", "terminal-after-recovery"],
+    );
+    let accepted = runtime
+        .restore_worker_operation_scoped(&scope, &worker_ref, request.clone())
+        .unwrap();
+    assert_eq!(accepted.state, server_api::WorkerRestoreState::Accepted);
+    assert_eq!(
+        runtime
+            .restore_worker_operation_scoped(&scope, &worker_ref, request.clone())
+            .unwrap(),
+        accepted
+    );
+    assert_eq!(restore_owner_count(&runtime, &worker_ref), 1);
+    let mut stale = request;
+    stale.request_id = "terminal-fresh-stale-intent".into();
+    assert!(matches!(
+        runtime.restore_worker_operation_scoped(&scope, &worker_ref, stale),
+        Err(crate::error::RuntimeError::RestoreObservationConflict { .. })
+    ));
+    assert_single_publication(&fixture, 1);
+    assert_eq!(
+        script.call_count.load(Ordering::SeqCst),
+        0,
+        "receipt recovery cannot trigger input or a model call"
+    );
+    // This transport was paired before the failed receipt write. Exercise its
+    // normal log-entry receiver after autonomous settlement, not a new attach.
+    transport
+        .dispatch(Method::submit_text(
+            "terminal-ordinary-input",
+            "terminal-ordinary-input",
+        ))
+        .unwrap();
+    fixture
+        .backend
+        .run_on_adapter_runtime(async move {
+            let mut transport = transport;
+            let (mut admitted, mut committed, mut text, mut ended) = (false, false, false, false);
+            tokio::time::timeout(WAIT, async {
+                while !(admitted && committed && text && ended) {
+                    match transport.events.recv().await.unwrap() {
+                        Event::SubmissionAccepted {
+                            submission_request_id,
+                            ..
+                        } if submission_request_id == "terminal-ordinary-input" => admitted = true,
+                        Event::UserMessage { segments, .. } => {
+                            assert_eq!(
+                                segments,
+                                vec![protocol::Segment::text("terminal-ordinary-input")]
+                            );
+                            assert!(!committed, "paired log receiver duplicated input");
+                            committed = true;
+                        }
+                        Event::TextDone { text: output } => {
+                            assert_eq!(output, "hello from worker");
+                            text = true;
+                        }
+                        Event::RunEnd { .. } => ended = true,
+                        Event::SubmissionRejected { .. } | Event::Error { .. } => {
+                            panic!("ordinary input failed after receipt recovery")
+                        }
+                        _ => {}
+                    }
+                }
+            })
+            .await
+            .map_err(|_| {
+                "ordinary input failed to settle on original paired transport".to_string()
+            })?;
+            drop(transport);
+            Ok(())
+        })
+        .unwrap();
+    assert_eq!(script.call_count.load(Ordering::SeqCst), 1);
+    assert_eq!(fixture.entries().iter().filter(|entry| matches!(entry,
+        LogEntry::AnnotatedUserInput { segments, .. } if segments == &vec![protocol::Segment::text("terminal-ordinary-input")]
+    )).count(), 1);
+    assert_single_publication(&fixture, 1);
+    drop(execution);
+    drop(current);
+    runtime
+        .stop_worker_scoped(&scope, &worker_ref, None)
+        .unwrap();
+    assert!(old.protocol_is_closed());
+    fixture.assert_released();
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn published_controller_terminal_receipt_before_rename_autonomously_settles_without_reconstruction()
+{
+    published_terminal_receipt_fault(TerminalFault::BeforeRename);
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn published_controller_terminal_receipt_after_rename_is_canonically_reread_and_synced() {
+    published_terminal_receipt_fault(TerminalFault::AfterRename);
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn published_controller_terminal_receipt_sync_unknown_retains_owner_and_write_fence() {
+    published_terminal_receipt_fault(TerminalFault::SyncUnknown);
+}
+
+#[test]
+#[serial_test::serial(worker_allocation)]
+fn published_controller_terminal_receipt_read_unknown_retains_owner_and_write_fence() {
+    published_terminal_receipt_fault(TerminalFault::ReadUnknown);
+}

crates/worker-runtime/src/worker_source.rs

@@ -1,23 +1,32 @@
 use std::sync::Arc;
-use std::time::{Duration, SystemTime, UNIX_EPOCH};
+#[cfg(feature = "http-server")]
+use std::time::Duration;
+use std::time::{SystemTime, UNIX_EPOCH};
 
+#[cfg(feature = "http-server")]
 use crate::workspace_request::{
     RuntimeWorkspaceRequest, RuntimeWorkspaceRequestClient, RuntimeWorkspaceRequestError,
 };
+use worker::WorkspaceResponse;
+#[cfg(feature = "http-server")]
 use worker::{
     WorkspaceBinaryRequest, WorkspaceBinaryResponse, WorkspaceClient, WorkspaceClientError,
     WorkspacePromptCatalogResolution, WorkspacePromptProjection, WorkspaceRequest,
-    WorkspaceRequestMethod, WorkspaceResponse, WorkspaceServerOperation,
+    WorkspaceRequestMethod, WorkspaceServerOperation,
 };
 
 use crate::auth::{
     RuntimeAuthError, RuntimeIdentityMaterial, RuntimeWorkerMutationSourceSigner,
-    WORKER_REMOVE_PERMISSION, WORKSPACE_REQUEST_PERMISSION, WORKSPACE_WORKER_DISCOVERY_PERMISSION,
-    WorkerMutationActorKind, WorkerMutationOperation, WorkerMutationSourceClaims, new_token_id,
+    WORKER_REMOVE_PERMISSION, WorkerMutationActorKind, WorkerMutationOperation,
+    WorkerMutationSourceClaims, new_token_id,
 };
+#[cfg(feature = "http-server")]
+use crate::auth::{WORKSPACE_REQUEST_PERMISSION, WORKSPACE_WORKER_DISCOVERY_PERMISSION};
 use crate::runtime::RuntimeWorkspaceScope;
+#[cfg(feature = "http-server")]
 use crate::worker_backend::WorkspacePromptProjectionCache;
 
+#[cfg(feature = "http-server")]
 fn ticket_assignment_role_segment(role: server_api::TicketAssignmentRole) -> &'static str {
     match role {
         server_api::TicketAssignmentRole::Orchestrator => "orchestrator",
@@ -144,6 +153,7 @@ pub trait EmbeddedWorkerMutationDispatcher: Send + Sync {
 
 #[derive(Clone)]
 enum RuntimeWorkerMutationTransport {
+    #[cfg(feature = "http-server")]
     Remote {
         request_client: RuntimeWorkspaceRequestClient,
     },
@@ -161,6 +171,7 @@ pub struct RuntimeWorkerMutationForwarder {
 }
 
 impl RuntimeWorkerMutationForwarder {
+    #[cfg(feature = "http-server")]
     pub fn remote(
         identity: &RuntimeIdentityMaterial,
         scope: RuntimeWorkspaceScope,
@@ -206,6 +217,7 @@ impl RuntimeWorkerMutationForwarder {
             target_worker_id,
         )?;
         match (&self.transport, proof) {
+            #[cfg(feature = "http-server")]
             (
                 RuntimeWorkerMutationTransport::Remote { request_client },
                 RuntimeOwnedWorkerMutationProof::Remote(token),
@@ -231,6 +243,7 @@ impl RuntimeWorkerMutationForwarder {
     }
 }
 
+#[cfg(feature = "http-server")]
 struct RemoteWorkerRemoveHttpRequest {
     request_client: RuntimeWorkspaceRequestClient,
     source_worker_id: String,
@@ -240,6 +253,7 @@ struct RemoteWorkerRemoveHttpRequest {
     reason: String,
 }
 
+#[cfg(feature = "http-server")]
 fn execute_remote_worker_remove_http(
     request: RemoteWorkerRemoveHttpRequest,
 ) -> Result<WorkspaceResponse, RuntimeWorkerMutationForwardError> {
@@ -263,6 +277,7 @@ fn execute_remote_worker_remove_http(
     execute_remote_worker_remove_http_blocking(request)
 }
 
+#[cfg(feature = "http-server")]
 fn execute_remote_worker_remove_http_blocking(
     request: RemoteWorkerRemoveHttpRequest,
 ) -> Result<WorkspaceResponse, RuntimeWorkerMutationForwardError> {
@@ -315,6 +330,7 @@ fn execute_remote_worker_remove_http_blocking(
     }
 }
 
+#[cfg(feature = "http-server")]
 #[derive(Clone)]
 pub struct RuntimeOwnedWorkspaceClient {
     workspace_id: String,
@@ -325,6 +341,7 @@ pub struct RuntimeOwnedWorkspaceClient {
     prompt_projection_cache: Option<Arc<WorkspacePromptProjectionCache>>,
 }
 
+#[cfg(feature = "http-server")]
 impl RuntimeOwnedWorkspaceClient {
     pub fn new(
         workspace_id: impl Into<String>,
@@ -495,6 +512,7 @@ impl RuntimeOwnedWorkspaceClient {
     }
 }
 
+#[cfg(feature = "http-server")]
 impl std::fmt::Debug for RuntimeOwnedWorkspaceClient {
     fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         formatter
@@ -510,6 +528,7 @@ impl std::fmt::Debug for RuntimeOwnedWorkspaceClient {
     }
 }
 
+#[cfg(feature = "http-server")]
 fn generated_workspace_response<T: serde::Serialize>(
     response: Result<T, server_api::client_support::ClientError<server_api::RepositoryApiError>>,
 ) -> Result<WorkspaceResponse, WorkspaceClientError> {
@@ -530,6 +549,7 @@ fn generated_workspace_response<T: serde::Serialize>(
     }
 }
 
+#[cfg(feature = "http-server")]
 impl WorkspaceClient for RuntimeOwnedWorkspaceClient {
     fn workspace_id(&self) -> Option<&str> {
         Some(self.request_client.workspace_id())
@@ -929,12 +949,17 @@ fn unix_now_seconds() -> u64 {
 #[cfg(test)]
 mod tests {
     use super::*;
+    #[cfg(feature = "http-server")]
     use crate::auth::{
-        RUNTIME_REQUEST_SOURCE_PROOF_HEADER, WorkerMutationSourceExpectation,
-        decode_runtime_request_source_claims, decode_worker_mutation_source_claims,
-        request_body_digest, verify_worker_mutation_source_proof,
+        RUNTIME_REQUEST_SOURCE_PROOF_HEADER, decode_runtime_request_source_claims,
+        request_body_digest,
+    };
+    use crate::auth::{
+        WorkerMutationSourceExpectation, decode_worker_mutation_source_claims,
+        verify_worker_mutation_source_proof,
     };
 
+    #[cfg(feature = "http-server")]
     fn binary_loopback(response: Vec<u8>) -> (String, std::thread::JoinHandle<(String, Vec<u8>)>) {
         use std::io::{Read, Write};
         let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap();
@@ -972,6 +997,7 @@ mod tests {
         (base_url, server)
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn generated_ticket_assignment_transport_signs_exact_authoring_and_bounded_read_subjects() {
         use crate::auth::{RuntimeRequestSourceExpectation, verify_runtime_request_source};
@@ -1104,6 +1130,7 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn binary_roundtrip_signs_exact_query_method_worker_and_raw_body() {
         use crate::auth::{RuntimeRequestSourceExpectation, verify_runtime_request_source};
@@ -1194,6 +1221,7 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn binary_responses_enforce_declared_and_streamed_limits_with_fixed_errors() {
         for response in [
@@ -1213,6 +1241,7 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn binary_redirect_returns_status_without_forwarding_authentication() {
         let (base_url, server) = binary_loopback(b"HTTP/1.1 307 Temporary Redirect\r\nLocation: http://127.0.0.1:1/private\r\nContent-Length: 0\r\nConnection: close\r\n\r\n".to_vec());
@@ -1231,6 +1260,7 @@ mod tests {
         assert!(result.body.is_empty());
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn binary_invalid_or_normalized_targets_fail_with_fixed_errors_before_http() {
         let client = RuntimeOwnedWorkspaceClient::new(
@@ -1277,6 +1307,7 @@ mod tests {
         );
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn current_prompt_projection_uses_the_shared_runtime_cache_without_http() {
         let cache = Arc::new(WorkspacePromptProjectionCache::default());
@@ -1312,6 +1343,7 @@ mod tests {
         assert!(Arc::ptr_eq(&projection.catalog, &second.catalog));
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn prompt_projection_cache_miss_fetches_content_identity() {
         use std::io::{Read, Write};
@@ -1356,6 +1388,7 @@ mod tests {
         assert_eq!(fetched.projection.source_digest, "source-3");
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn concurrent_prompt_projection_miss_fetches_once_and_shares_catalog() {
         use std::io::{Read, Write};
@@ -1427,6 +1460,7 @@ mod tests {
         );
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn current_prompt_projection_rejects_cross_workspace_response() {
         use std::io::{Read, Write};
@@ -1473,6 +1507,7 @@ mod tests {
         assert!(error.to_string().contains("scope mismatch"));
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn ordinary_workspace_forwarding_stamps_runtime_identity_and_signs_path_and_query() {
         use std::io::{Read, Write};
@@ -1531,6 +1566,7 @@ mod tests {
         );
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn workspace_worker_discovery_signs_dedicated_permission_and_encoded_query() {
         use std::io::{Read, Write};
@@ -1652,6 +1688,7 @@ mod tests {
         );
     }
 
+    #[cfg(feature = "http-server")]
     #[tokio::test(flavor = "multi_thread")]
     async fn remote_forwarder_is_safe_in_async_runtime_and_stamps_signed_proof() {
         use std::io::{Read, Write};
@@ -1815,6 +1852,7 @@ mod tests {
         assert_eq!(reason, "retire obsolete Worker");
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn runtime_owned_workspace_client_returns_successful_response_without_request_timeout() {
         // Effective deadline clearing is exercised with a short inherited deadline in
@@ -1838,6 +1876,7 @@ mod tests {
         server.join().unwrap();
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn runtime_owned_workspace_client_honors_per_request_timeouts() {
         let (base_url, server) = delayed_workspace_response(Duration::from_millis(75));
@@ -1857,6 +1896,7 @@ mod tests {
         server.join().unwrap();
     }
 
+    #[cfg(feature = "http-server")]
     #[test]
     fn runtime_owned_workspace_client_reports_request_timeouts() {
         let (base_url, server) = delayed_workspace_response(Duration::from_millis(75));
@@ -1878,6 +1918,7 @@ mod tests {
         server.join().unwrap();
     }
 
+    #[cfg(feature = "http-server")]
     fn delayed_workspace_response(delay: Duration) -> (String, std::thread::JoinHandle<()>) {
         use std::io::{Read, Write};
         use std::net::TcpListener;

crates/worker-runtime/src/working_directory.rs

@@ -3676,6 +3676,7 @@ mod tests {
 
     #[test]
     fn bound_ssh_workdir_uses_attachment_scoped_agent_and_cleans_it_up() {
+        crate::restore_test_environment::require_restore_test_environment();
         let repo = create_clean_repo();
         // ssh-agent and the broker receive relative socket paths in this owned
         // fixture tree, so the test does not depend on checkout/TMPDIR length.

crates/worker/src/bootstrap.rs

@@ -7,6 +7,7 @@ use workdir::WorkdirSessionHandle;
 
 use crate::PromptCatalogSource;
 use crate::controller::{ShutdownReceiver, WorkerController, WorkerHandle};
+pub use crate::worker::{PreparedRestoredWorker, RestoreCandidateWorkdirCleanup};
 use crate::worker::{Worker, WorkerError, WorkerFilesystemAuthority, WorkerWorkspaceContext};
 use manifest::WorkerManifest;
 
@@ -161,6 +162,38 @@ where
         Ok(PreparedWorker::new(worker, self.layout))
     }
 
+    /// Reconstruct a saved materialized Worker without allocation, attribution
+    /// repair, Feature installation, or controller exposure. The caller owns the
+    /// later controller layout and must activate before starting it.
+    pub async fn prepare_restored_candidate(
+        self,
+        worker_name: &str,
+    ) -> Result<PreparedRestoredWorker<St>, WorkerBootstrapError> {
+        let mut candidate =
+            Worker::prepare_restored_from_worker_metadata_with_context_and_model_client(
+                worker_name,
+                self.manifest,
+                self.store,
+                self.prompt_catalog,
+                self.workspace_context,
+                self.filesystem_authority,
+                self.model_client,
+            )
+            .await
+            .map_err(WorkerBootstrapError::Worker)?;
+        if let Some(workdir_session) = self.workdir_session {
+            candidate
+                .worker_mut()
+                .map_err(WorkerBootstrapError::Worker)?
+                .bind_single_workdir_session(Some(workdir_session));
+        }
+        candidate
+            .worker_mut()
+            .map_err(WorkerBootstrapError::Worker)?
+            .bind_subjektiv_host(self.subjektiv_host);
+        Ok(candidate)
+    }
+
     pub async fn prepare_restored(
         self,
         worker_name: &str,
@@ -243,6 +276,7 @@ where
     St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
 {
     let cleanup_sessions = worker.workdir_sessions();
+    let factory_owned_cleanup = worker.startup_feature_cleanup_handle().is_some();
     let controller = match layout {
         WorkerBootstrapLayout::Direct { runtime_base } => {
             WorkerController::spawn_owned(worker, &runtime_base).await
@@ -259,7 +293,10 @@ where
             controller_task,
         }),
         Err(source) => {
-            let cleanup_failed = cleanup_sessions.close_all().await.is_err();
+            // Factory-owned task scopes must settle before sessions close.
+            // Cleanup remains unconfirmed here; preserve the startup source.
+            let cleanup_failed =
+                factory_owned_cleanup || cleanup_sessions.close_all().await.is_err();
             Err(WorkerBootstrapError::Controller {
                 source,
                 cleanup_failed,
@@ -267,3 +304,229 @@ where
         }
     }
 }
+
+/// Host bindings for one unpublished restored Controller. These do not expose
+/// Worker execution or permit arbitrary mutation during preparation.
+pub struct RestoredControllerBindings {
+    pub backend_job: Option<crate::BackendJobExecutionBinding>,
+    pub workdir_sessions: std::sync::Arc<workdir::WorkdirSessionRouter>,
+    pub observation_provider:
+        Option<std::sync::Arc<dyn crate::feature::builtin::WorkerObservationProvider>>,
+    pub subjektiv_attached: bool,
+}
+
+/// The same real receiver, snapshot sources and paired subscription retained
+/// privately across the Runtime's durable restore decision.
+pub struct UnpublishedRestoredController<St: Store> {
+    initialized: Option<crate::controller::InitializedWorkerController<Box<dyn LlmClient>, St>>,
+    activation: Option<crate::worker::RestoredWorkerActivation>,
+    streams: Option<crate::ipc::protocol_session::WorkerProtocolSessionStreams>,
+    subjektiv_attached: bool,
+    effects_applied: bool,
+    activation_failed: bool,
+    closing: bool,
+}
+
+pub struct ActivatedRestoredController {
+    pub controller: BootstrappedWorker,
+    pub streams: crate::ipc::protocol_session::WorkerProtocolSessionStreams,
+}
+
+impl<St> PreparedRestoredWorker<St>
+where
+    St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+{
+    pub fn workspace_client_handle(
+        &mut self,
+    ) -> Result<std::sync::Arc<dyn crate::WorkspaceClient>, WorkerError> {
+        Ok(self.worker_mut()?.workspace_client_handle())
+    }
+
+    pub fn retain_startup_feature_cleanup(
+        &mut self,
+    ) -> Result<crate::WorkerStartupFeatureCleanup, WorkerError> {
+        Ok(self.worker_mut()?.retain_startup_feature_cleanup())
+    }
+
+    /// Initialize the real Controller without allocation, canonical restore
+    /// writes, attribution changes, registry reclamation or input consumption.
+    /// MCP discovery and Feature tasks belong to the retained startup cleanup.
+    pub async fn prepare_runtime_controller(
+        mut self,
+        run_dir: &std::path::Path,
+        bindings: RestoredControllerBindings,
+    ) -> Result<UnpublishedRestoredController<St>, WorkerBootstrapError> {
+        let worker = self.worker_mut().map_err(WorkerBootstrapError::Worker)?;
+        if let Some(binding) = bindings.backend_job {
+            worker.bind_backend_job(binding).map_err(|error| {
+                WorkerBootstrapError::Worker(WorkerError::FeatureLifecycle(error.into()))
+            })?;
+        }
+        worker
+            .workdir_sessions()
+            .close_all()
+            .await
+            .map_err(|error| {
+                WorkerBootstrapError::Worker(WorkerError::FeatureLifecycle(error.to_string()))
+            })?;
+        worker.bind_workdir_sessions(bindings.workdir_sessions);
+        worker.bind_worker_observation_provider(bindings.observation_provider);
+        let flow_enabled = worker.manifest().feature.flow.enabled;
+        if flow_enabled {
+            let report = worker
+                .install_runtime_flow_transition_feature()
+                .map_err(|error| {
+                    WorkerBootstrapError::Worker(WorkerError::FeatureLifecycle(error))
+                })?;
+            if report.reports.iter().any(|report| !report.installed) {
+                return Err(WorkerBootstrapError::Worker(WorkerError::FeatureLifecycle(
+                    "Flow transition installation refused".into(),
+                )));
+            }
+        }
+        let (worker, activation) = self
+            .into_controller_parts()
+            .map_err(WorkerBootstrapError::Worker)?;
+        let initialized = WorkerController::initialize_restored(worker, run_dir)
+            .await
+            .map_err(|source| WorkerBootstrapError::Controller {
+                source,
+                cleanup_failed: true,
+            })?;
+        if flow_enabled {
+            initialized.handle.shared_state.enable_flow_transition();
+        }
+        let streams = {
+            let _guard = initialized.handle.protocol_snapshot_publish_guard();
+            crate::ipc::protocol_session::subscribe_worker_protocol_session(&initialized.handle)
+        };
+        Ok(UnpublishedRestoredController {
+            initialized: Some(initialized),
+            activation: Some(activation),
+            streams: Some(streams),
+            subjektiv_attached: bindings.subjektiv_attached,
+            effects_applied: false,
+            activation_failed: false,
+            closing: false,
+        })
+    }
+}
+
+impl<St> UnpublishedRestoredController<St>
+where
+    St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+{
+    pub fn snapshot(&self) -> &protocol::Event {
+        &self
+            .streams
+            .as_ref()
+            .expect("unpublished restore streams")
+            .snapshot_event
+    }
+
+    /// Only publication failures after confirmed effects can retry on this
+    /// exact owner. Partial activation must be reconciled, never blindly rerun.
+    pub fn can_retry_publication(&self) -> bool {
+        self.initialized.is_some()
+            && self.effects_applied
+            && !self.activation_failed
+            && !self.closing
+    }
+
+    /// Apply canonical restore effects after the host's durable decision, while
+    /// retaining the same private Controller for subsequent publication.
+    pub async fn apply_effects(&mut self) -> Result<(), WorkerBootstrapError> {
+        let initialized = self
+            .initialized
+            .as_mut()
+            .ok_or_else(|| WorkerBootstrapError::Worker(WorkerError::RestoreCandidateClosing))?;
+        if self.activation_failed || self.closing {
+            return Err(WorkerBootstrapError::Worker(
+                WorkerError::RestoreCandidateClosing,
+            ));
+        }
+        // Preparation's read-only authorization is not a publication grant.
+        // Revalidate current Host authority after commit, before effects/input,
+        // including a publication retry on this same private Controller.
+        initialized
+            .worker
+            .validate_subjektiv_body_publication()
+            .map_err(WorkerBootstrapError::Worker)?;
+        if !self.effects_applied {
+            self.activation_failed = true;
+            self.activation
+                .as_ref()
+                .expect("restore activation")
+                .activate(&mut initialized.worker)
+                .await
+                .map_err(WorkerBootstrapError::Worker)?;
+            initialized
+                .worker
+                .finalize_subjektiv_session_attribution(
+                    self.subjektiv_attached,
+                    crate::SubjektivSessionAttributionLifecycle::RestoredSession,
+                )
+                .map_err(WorkerBootstrapError::Worker)?;
+            self.effects_applied = true;
+            self.activation_failed = false;
+            self.activation = None;
+        }
+        Ok(())
+    }
+
+    pub fn allocation_cleanup(
+        &self,
+    ) -> Option<crate::runtime::worker_allocation::ScopeAllocationCleanup> {
+        self.initialized
+            .as_ref()
+            .and_then(|initialized| initialized.worker.restore_allocation_cleanup_handle())
+    }
+
+    pub async fn activate(&mut self) -> Result<ActivatedRestoredController, WorkerBootstrapError> {
+        self.apply_effects().await?;
+        let initialized = self
+            .initialized
+            .as_mut()
+            .expect("same initialized Controller");
+        let publication = initialized.prepare_publication().await.map_err(|source| {
+            WorkerBootstrapError::Controller {
+                source,
+                cleanup_failed: true,
+            }
+        })?;
+        let initialized = self
+            .initialized
+            .take()
+            .expect("same initialized Controller");
+        let (handle, shutdown, controller_task) = initialized.start(publication);
+        Ok(ActivatedRestoredController {
+            controller: BootstrappedWorker {
+                handle,
+                shutdown,
+                controller_task,
+            },
+            streams: self.streams.take().expect("same paired subscription"),
+        })
+    }
+
+    pub async fn discard(&mut self) -> Result<(), WorkerError> {
+        self.closing = true;
+        if let Some(initialized) = self.initialized.as_mut() {
+            initialized
+                .worker
+                .retain_startup_feature_cleanup()
+                .shutdown()
+                .await
+                .map_err(|error| WorkerError::FeatureLifecycle(error.to_string()))?;
+            initialized
+                .worker
+                .workdir_sessions()
+                .close_all()
+                .await
+                .map_err(|error| WorkerError::FeatureLifecycle(error.to_string()))?;
+        }
+        self.streams = None;
+        self.initialized = None;
+        Ok(())
+    }
+}

crates/worker/src/controller.rs

@@ -21,7 +21,7 @@ use crate::shutdown_after_idle::{
     ShutdownAfterIdleRequest, TicketIntakeReadyShutdownHook, is_ticket_intake_role,
     take_shutdown_request_after_status,
 };
-use crate::spawn::registry::SpawnedWorkerRegistry;
+use crate::spawn::registry::{PreparedSpawnedWorkerRegistryLoad, SpawnedWorkerRegistry};
 use crate::spawn::tool::sub_worker_spawn_tool;
 use crate::worker::{
     PendingSubmissionHandle, SystemItemCommitter, Worker, WorkerError, WorkerRunResult,
@@ -654,7 +654,56 @@ pub type ShutdownReceiver = oneshot::Receiver<()>;
 
 pub struct WorkerController;
 
+// Own the real command receiver and all snapshot sources before starting any
+// protocol listener or controller loop. Fresh startup may materialize the head;
+// Runtime restore requires an already materialized source and defers canonical
+// restore effects until its durable decision. Tool discovery can own external
+// resources, so private initialization is not synonymous with side-effect-free.
+pub(crate) struct InitializedWorkerController<C: LlmClient, St: Store> {
+    pub(crate) worker: Worker<C, St>,
+    pub(crate) handle: WorkerHandle,
+    method_rx: mpsc::Receiver<Method>,
+    shutdown_tx: oneshot::Sender<()>,
+    shutdown_rx: ShutdownReceiver,
+    registry_reclaim: PreparedSpawnedWorkerRegistryLoad<St>,
+    shutdown_after_idle: ShutdownAfterIdleRequest,
+    command_events: Option<PreparedWorkdirCommandEvents>,
+}
+
+pub(crate) struct PreparedControllerPublication {
+    command_observer: Option<tokio::task::JoinHandle<()>>,
+}
+
 impl WorkerController {
+    /// Explicit recovery of an unmaterialized Session head. This installs the
+    /// saved Feature composition needed to produce the seed, but creates no
+    /// Controller, input channel, socket, observer, or queue consumer. The caller
+    /// owns the same operation's Feature/session/allocation cleanup on every exit.
+    pub async fn recover_pending_head<C, St>(
+        worker: &mut Worker<C, St>,
+        runtime_base: &Path,
+    ) -> std::io::Result<()>
+    where
+        C: LlmClient + Clone + 'static,
+        St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+    {
+        if worker.needs_initial_session_head_materialization() {
+            worker
+                .validate_subjektiv_body_publication()
+                .map_err(std::io::Error::other)?;
+            let registry = SpawnedWorkerRegistry::new_internal(
+                worker.manifest().worker.name.clone(),
+                worker.scope().clone(),
+            );
+            register_worker_tools(worker, runtime_base.to_path_buf(), registry, None, None).await?;
+            worker
+                .materialize_durable_session_head()
+                .await
+                .map_err(std::io::Error::other)?;
+        }
+        Ok(())
+    }
+
     pub async fn spawn<C, St>(
         worker: Worker<C, St>,
         runtime_base: &Path,
@@ -734,8 +783,12 @@ impl WorkerController {
         St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
     {
         let sessions = worker.workdir_sessions();
+        let factory_owned_cleanup = worker.startup_feature_cleanup_handle().is_some();
         let result = Self::spawn_initialized(worker, runtime_base, runtime_run).await;
+        // An opted-in factory must quiesce its captured Feature tasks before
+        // closing their Workdir sessions; this wrapper cannot confirm that order.
         if result.is_err()
+            && !factory_owned_cleanup
             && let Err(error) = sessions.close_all().await
         {
             tracing::warn!(%error, "Workdir sessions close after controller startup failure failed");
@@ -744,7 +797,7 @@ impl WorkerController {
     }
 
     async fn spawn_initialized<C, St>(
-        mut worker: Worker<C, St>,
+        worker: Worker<C, St>,
         runtime_base: &Path,
         runtime_run: Option<&Path>,
     ) -> Result<(WorkerHandle, ShutdownReceiver, tokio::task::JoinHandle<()>), std::io::Error>
@@ -752,6 +805,52 @@ impl WorkerController {
         C: LlmClient + Clone + 'static,
         St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
     {
+        Self::initialize(worker, runtime_base, runtime_run)
+            .await?
+            .publish()
+            .await
+    }
+
+    async fn initialize<C, St>(
+        worker: Worker<C, St>,
+        runtime_base: &Path,
+        runtime_run: Option<&Path>,
+    ) -> Result<InitializedWorkerController<C, St>, std::io::Error>
+    where
+        C: LlmClient + Clone + 'static,
+        St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+    {
+        Self::initialize_inner(worker, runtime_base, runtime_run, true).await
+    }
+
+    pub(crate) async fn initialize_restored<St>(
+        worker: Worker<Box<dyn LlmClient>, St>,
+        run_dir: &Path,
+    ) -> std::io::Result<InitializedWorkerController<Box<dyn LlmClient>, St>>
+    where
+        St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+    {
+        let parent = run_dir
+            .parent()
+            .ok_or_else(|| std::io::Error::other("run path has no parent"))?;
+        Self::initialize_inner(worker, parent, Some(run_dir), false).await
+    }
+
+    async fn initialize_inner<C, St>(
+        mut worker: Worker<C, St>,
+        runtime_base: &Path,
+        runtime_run: Option<&Path>,
+        canonical_startup: bool,
+    ) -> Result<InitializedWorkerController<C, St>, std::io::Error>
+    where
+        C: LlmClient + Clone + 'static,
+        St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+    {
+        if !canonical_startup && worker.needs_initial_session_head_materialization() {
+            return Err(std::io::Error::other(
+                "restore source requires operation-owned head recovery",
+            ));
+        }
         worker
             .validate_subjektiv_body_publication()
             .map_err(std::io::Error::other)?;
@@ -773,21 +872,13 @@ impl WorkerController {
         });
 
         let spawner_name = worker.manifest().worker.name.clone();
-        let loaded_registry = SpawnedWorkerRegistry::load_from_worker_state_with_reclaim(
+        let registry_reclaim = SpawnedWorkerRegistry::prepare_from_worker_state_with_reclaim(
             runtime_dir.clone(),
             worker.store().clone(),
             spawner_name.clone(),
             Some(worker.scope().clone()),
-        )
-        .await?;
-        let reclaimed_unreachable = loaded_registry.reclaimed_unreachable;
-        let spawned_registry = loaded_registry.registry;
-        if reclaimed_unreachable {
-            worker.push_notify(
-                "Restored Worker state contained unreachable delegated child Workers; their delegated write scopes were reclaimed before resume."
-                    .to_string(),
-            );
-        }
+        )?;
+        let spawned_registry = registry_reclaim.registry.clone();
 
         // Hand the alerter to the Worker so internal operations (compaction,
         // AGENTS.md ingestion during the first turn) can emit user-facing
@@ -878,7 +969,6 @@ impl WorkerController {
         if let Some(fs_for_view) = fs_for_view.as_ref() {
             shared_state.set_fs_view(crate::fs_view::WorkerFsView::new(fs_for_view.clone()));
         }
-
         let artifact_store: Arc<dyn Store> = Arc::new(worker.store().clone());
         let session_id = worker.session_id();
         let pending_activations = worker.pending_activation_state();
@@ -903,10 +993,71 @@ impl WorkerController {
             pending_activations,
         };
 
-        // Start the observer only after fallible head/recovery initialization.
-        let command_observer = fs_for_view
+        let command_events = fs_for_view
             .as_ref()
-            .and_then(|session| wire_workdir_command_events(session, &in_flight));
+            .map(|session| PreparedWorkdirCommandEvents::new(session, &handle.in_flight));
+        Ok(InitializedWorkerController {
+            worker,
+            handle,
+            method_rx,
+            shutdown_tx,
+            shutdown_rx,
+            registry_reclaim,
+            shutdown_after_idle,
+            command_events,
+        })
+    }
+}
+
+impl<C, St> InitializedWorkerController<C, St>
+where
+    C: LlmClient + Clone + 'static,
+    St: Store + WorkerMetadataStore + Clone + Send + Sync + 'static,
+{
+    async fn publish(
+        mut self,
+    ) -> Result<(WorkerHandle, ShutdownReceiver, tokio::task::JoinHandle<()>), std::io::Error> {
+        let publication = self.prepare_publication().await?;
+        Ok(self.start(publication))
+    }
+
+    pub(crate) async fn prepare_publication(
+        &mut self,
+    ) -> std::io::Result<PreparedControllerPublication> {
+        let loaded_registry = self.registry_reclaim.apply().await?;
+        debug_assert!(Arc::ptr_eq(
+            &loaded_registry.registry,
+            &self.handle.spawned_registry
+        ));
+        if loaded_registry.reclaimed_unreachable {
+            self.worker.push_notify("Restored Worker state contained unreachable delegated child Workers; their delegated write scopes were reclaimed before resume.".to_string());
+        }
+        let command_observer = self
+            .command_events
+            .take()
+            .and_then(PreparedWorkdirCommandEvents::start);
+        Ok(PreparedControllerPublication { command_observer })
+    }
+
+    pub(crate) fn start(
+        self,
+        publication: PreparedControllerPublication,
+    ) -> (WorkerHandle, ShutdownReceiver, tokio::task::JoinHandle<()>) {
+        let Self {
+            mut worker,
+            handle,
+            method_rx,
+            shutdown_tx,
+            shutdown_rx,
+            shutdown_after_idle,
+            ..
+        } = self;
+        let PreparedControllerPublication { command_observer } = publication;
+        let runtime_dir = handle.runtime_dir.clone();
+        let shared_state = handle.shared_state.clone();
+        let working_event_tx = handle.working_event_tx.clone();
+        let spawned_registry = handle.spawned_registry.clone();
+
         // === 5. controller_loop ===
         // Clone cancel sender and notification buffer before moving worker
         // into the controller task so the in-flight turn can be reached
@@ -915,6 +1066,7 @@ impl WorkerController {
         let pause_tx = worker.engine_mut().pause_sender();
         let notify_buffer = worker.notify_buffer_handle();
 
+        worker.finish_startup_feature_cleanup_capture();
         let controller_task = tokio::spawn(controller_loop(
             worker,
             method_rx,
@@ -930,7 +1082,7 @@ impl WorkerController {
             command_observer,
         ));
 
-        Ok((handle, shutdown_rx, controller_task))
+        (handle, shutdown_rx, controller_task)
     }
 }
 
@@ -938,30 +1090,51 @@ pub(crate) fn wire_workdir_command_events(
     session: &Arc<dyn WorkdirSession>,
     in_flight: &InFlightEvents,
 ) -> Option<tokio::task::JoinHandle<()>> {
-    in_flight.replace_command_snapshot(protocol_command_snapshots(session.as_ref()));
-    let Some(mut events) = session.subscribe_command_events() else {
-        return None;
-    };
-    // Keep only a weak reference in the observer task. Holding the session
-    // strongly here would keep its broadcast sender alive forever and prevent
-    // the receiver from observing closure during Worker teardown.
-    let session = Arc::downgrade(session);
-    let in_flight = in_flight.clone();
-    Some(tokio::spawn(async move {
-        loop {
-            match events.recv().await {
-                Ok(event) => in_flight.publish_command_event(protocol_command_event(event)),
-                Err(broadcast::error::RecvError::Lagged(_)) => {
-                    let Some(session) = session.upgrade() else {
-                        break;
-                    };
-                    in_flight
-                        .replace_command_snapshot(protocol_command_snapshots(session.as_ref()));
-                }
-                Err(broadcast::error::RecvError::Closed) => break,
-            }
+    PreparedWorkdirCommandEvents::new(session, in_flight).start()
+}
+
+struct PreparedWorkdirCommandEvents {
+    session: std::sync::Weak<dyn WorkdirSession>,
+    in_flight: InFlightEvents,
+    events: Option<broadcast::Receiver<WorkdirCommandEvent>>,
+}
+
+impl PreparedWorkdirCommandEvents {
+    fn new(session: &Arc<dyn WorkdirSession>, in_flight: &InFlightEvents) -> Self {
+        // Subscribe before reading command state; hold the receiver privately
+        // across the durable decision instead of creating a publication gap.
+        let events = session.subscribe_command_events();
+        in_flight.replace_command_snapshot(protocol_command_snapshots(session.as_ref()));
+        Self {
+            session: Arc::downgrade(session),
+            in_flight: in_flight.clone(),
+            events,
         }
-    }))
+    }
+
+    fn start(self) -> Option<tokio::task::JoinHandle<()>> {
+        let Self {
+            session,
+            in_flight,
+            events,
+        } = self;
+        let mut events = events?;
+        Some(tokio::spawn(async move {
+            loop {
+                match events.recv().await {
+                    Ok(event) => in_flight.publish_command_event(protocol_command_event(event)),
+                    Err(broadcast::error::RecvError::Lagged(_)) => {
+                        let Some(session) = session.upgrade() else {
+                            break;
+                        };
+                        in_flight
+                            .replace_command_snapshot(protocol_command_snapshots(session.as_ref()));
+                    }
+                    Err(broadcast::error::RecvError::Closed) => break,
+                }
+            }
+        }))
+    }
 }
 
 fn protocol_command_snapshots(session: &dyn WorkdirSession) -> Vec<ProtocolCommandSnapshot> {
@@ -1687,9 +1860,17 @@ where
             .add_module(crate::feature::builtin::orchestration::orchestration_feature());
     }
     if let Some(workspace_root) = local_workspace_root.as_ref() {
-        if let Some(module) =
+        let module = if let Some(cleanup) = worker.startup_feature_cleanup_handle() {
+            crate::feature::mcp::discover_stdio_tool_feature_with_cleanup(
+                &mcp_config,
+                workspace_root,
+                &cleanup.mcp_discovery(),
+            )
+            .await?
+        } else {
             crate::feature::mcp::discover_stdio_tool_feature(&mcp_config, workspace_root).await
-        {
+        };
+        if let Some(module) = module {
             feature_registry = feature_registry.with_module(module);
         }
     }
@@ -3750,6 +3931,159 @@ mod tests {
             }
         }
 
+        // Component connection: real Worker/Store, initialized endpoint, paired
+        // protocol subscription and Controller loop. The model boundary is an
+        // empty non-network script; this test must not request any generation.
+        #[tokio::test]
+        async fn initialized_controller_publishes_the_same_endpoint_and_subscription() {
+            let root = TempDir::new().unwrap();
+            let manifest = manifest::WorkerManifest::from_toml(
+                "[worker]\nname = 'p'\n[model]\nscheme = 'anthropic'\nmodel_id = 'test'\n[engine]\n[scope]\n",
+            )
+            .unwrap();
+            let client = ScriptedClient::new(vec![]);
+            let requests = client.requests.clone();
+            let engine =
+                Engine::<_, agen::state::Mutable, crate::SessionHistoryMetadata>::new_annotated(
+                    client,
+                );
+            let store = session_store::CombinedStore::new(
+                session_store::FsStore::new(root.path().join("sessions")).unwrap(),
+                session_store::FsWorkerStore::new(root.path().join("workers")).unwrap(),
+            );
+            let mut worker = Worker::new(
+                manifest,
+                engine,
+                store,
+                crate::worker::WorkerWorkspaceContext::local_filesystem(None),
+                crate::worker::WorkerFilesystemAuthority::None,
+                manifest::Scope::empty(),
+            )
+            .await
+            .unwrap();
+            worker.set_system_prompt_template(
+                crate::SystemPromptTemplate::parse(
+                    "default",
+                    crate::PromptCatalogSource::builtins_only(),
+                )
+                .unwrap(),
+            );
+            let prepared = WorkerController::initialize(worker, root.path(), None)
+                .await
+                .unwrap();
+            let endpoint = prepared.handle.clone();
+            let mut streams =
+                crate::ipc::protocol_session::subscribe_worker_protocol_session(&endpoint);
+            let entries = endpoint.committed_entries();
+            assert!(matches!(
+                &streams.snapshot_event,
+                Event::Snapshot { greeting, state, session, .. }
+                    if !greeting.tools.is_empty()
+                        && state.state == WorkerState::Idle
+                        && session == &session_store::public_snapshot::project_current_session_snapshot(&entries)
+            ));
+            assert!(!endpoint.protocol_is_closed());
+            assert!(!endpoint.runtime_dir.path().join("sock").exists());
+            assert!(!endpoint.runtime_dir.path().join("status.json").exists());
+            assert!(!endpoint.runtime_dir.path().join("manifest.toml").exists());
+            assert!(matches!(
+                streams.events.try_recv(),
+                Err(broadcast::error::TryRecvError::Empty)
+            ));
+            assert!(requests.lock().unwrap().is_empty());
+
+            let (handle, shutdown, task) = prepared.publish().await.unwrap();
+            assert!(handle.same_controller(&endpoint));
+            assert!(!handle.runtime_dir.path().join("sock").exists());
+            assert!(!handle.runtime_dir.path().join("status.json").exists());
+            assert!(!handle.runtime_dir.path().join("manifest.toml").exists());
+            handle.send(Method::ListPendingSubmissions).await.unwrap();
+            let pending = tokio::time::timeout(Duration::from_secs(5), async {
+                loop {
+                    if let Event::PendingSubmissionsChanged { pending } =
+                        streams.events.recv().await.unwrap()
+                    {
+                        break pending;
+                    }
+                }
+            })
+            .await
+            .unwrap();
+            assert_eq!(pending, protocol::PendingSubmissionsSnapshot::default());
+            assert_eq!(
+                serde_json::to_value(handle.committed_entries()).unwrap(),
+                serde_json::to_value(entries).unwrap(),
+            );
+            handle
+                .send(Method::Shutdown {
+                    command: WorkerCommandEnvelope { command_id: 1 },
+                })
+                .await
+                .unwrap();
+            tokio::time::timeout(Duration::from_secs(5), shutdown)
+                .await
+                .unwrap()
+                .unwrap();
+            tokio::time::timeout(Duration::from_secs(5), task)
+                .await
+                .unwrap()
+                .unwrap();
+            assert!(endpoint.protocol_is_closed());
+            assert!(!endpoint.runtime_dir.path().join("sock").exists());
+            assert!(requests.lock().unwrap().is_empty());
+        }
+
+        #[tokio::test]
+        async fn tool_initialization_failure_does_not_reclaim_saved_children() {
+            let root = TempDir::new().unwrap();
+            let mut manifest = manifest::WorkerManifest::from_toml(
+                "[worker]\nname = 'p'\n[model]\nscheme = 'anthropic'\nmodel_id = 'test'\n[engine]\n[scope]\n",
+            ).unwrap();
+            manifest.feature.ticket.enabled = true;
+            let client = ScriptedClient::new(vec![]);
+            let requests = client.requests.clone();
+            let store = session_store::CombinedStore::new(
+                session_store::FsStore::new(root.path().join("sessions")).unwrap(),
+                session_store::FsWorkerStore::new(root.path().join("workers")).unwrap(),
+            );
+            // New feature activation is refused by the real unavailable host;
+            // legacy child metadata must not be consumed before that refusal.
+            let worker = Worker::new(
+                manifest,
+                Engine::<_, agen::state::Mutable, crate::SessionHistoryMetadata>::new_annotated(
+                    client,
+                ),
+                store.clone(),
+                crate::worker::WorkerWorkspaceContext::no_workspace(),
+                crate::worker::WorkerFilesystemAuthority::None,
+                manifest::Scope::empty(),
+            )
+            .await
+            .unwrap();
+            let mut metadata = session_store::WorkerMetadata::new(
+                "p",
+                Some(session_store::WorkerActiveSegmentRef {
+                    session_id: worker.session_id(),
+                    segment_id: Some(worker.segment_id()),
+                }),
+            );
+            metadata
+                .spawned_children
+                .push(session_store::WorkerSpawnedChild {
+                    worker_name: "legacy".into(),
+                    socket_path: root.path().join("legacy.sock"),
+                    callback_address: root.path().join("p/sock"),
+                    scope_delegated: Vec::new(),
+                });
+            store.write(&metadata).unwrap();
+            let before = store.read_by_name("p").unwrap();
+            let result = WorkerController::initialize(worker, root.path(), None).await;
+            assert!(result.is_err());
+            assert_eq!(store.read_by_name("p").unwrap(), before);
+            assert!(!root.path().join("p/spawned_workers.json").exists());
+            assert!(requests.lock().unwrap().is_empty());
+        }
+
         #[derive(Debug)]
         struct DriveWorkspaceHost {
             available: bool,
@@ -4077,10 +4411,19 @@ mod tests {
         assert!(feature_install < materialize_gate);
         assert!(materialize_gate < materialize);
         assert!(materialize < exposure);
-        let observer = startup.find("let command_observer = fs_for_view").unwrap();
+        let observer_preparation = startup
+            .find("PreparedWorkdirCommandEvents::new(session")
+            .unwrap();
         let recover = startup.find(".recover_unfinished_compaction()").unwrap();
+        let reclaim = startup
+            .find("self.registry_reclaim.apply().await?")
+            .unwrap();
+        let observer_start = startup
+            .find(".and_then(PreparedWorkdirCommandEvents::start)")
+            .unwrap();
         let loop_start = startup.find("tokio::spawn(controller_loop(").unwrap();
-        assert!(recover < observer && observer < loop_start);
+        assert!(recover < observer_preparation && observer_preparation < reclaim);
+        assert!(reclaim < observer_start && observer_start < loop_start);
         assert!(
             startup[materialize_gate..materialize]
                 .contains("needs_initial_session_head_materialization")

crates/worker/src/feature/mcp.rs

@@ -8,6 +8,7 @@
 //! the regular registry decides Tool-name collisions.
 
 use std::collections::BTreeMap;
+use std::io;
 use std::path::Path;
 use std::sync::Arc;
 
@@ -25,6 +26,7 @@ use mcp::stdio::{
 };
 use serde::Serialize;
 use serde_json::{Map, Value};
+use tokio::sync::Mutex;
 
 use super::{
     FeatureDescriptor, FeatureDiagnostic, FeatureInstallContext, FeatureInstallError,
@@ -52,8 +54,88 @@ const MAX_LIST_ITEMS: usize = 128;
 const MAX_RESOURCE_CONTENTS: usize = 16;
 const MAX_PROMPT_MESSAGES: usize = 32;
 
+/// Owns discovery clients until process reaping is confirmed. Retain a clone in
+/// startup cleanup *before* awaiting discovery. Stop/cancel discovery before
+/// calling `shutdown`; the shared lock serializes discovery and cleanup, not
+/// cancellation of an in-flight discovery task.
+///
+/// Error or cancellation leaves the actual clients here for another shutdown
+/// attempt. Dropping the final owner only provides best-effort process killing.
+#[derive(Clone, Default)]
+pub struct McpDiscoveryCleanup {
+    clients: Arc<Mutex<Vec<McpStdioClient>>>,
+}
+
+impl McpDiscoveryCleanup {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Try every retained client once, removing only confirmed shutdowns. An
+    /// error is not a cleanup receipt; retry this same owner, not discovery.
+    pub async fn shutdown(&self) -> io::Result<()> {
+        let mut clients = self.clients.lock().await;
+        let mut first_error = None;
+        let mut index = 0;
+        while index < clients.len() {
+            match clients[index].shutdown().await {
+                Ok(_) => {
+                    clients.remove(index);
+                }
+                Err(err) => {
+                    first_error.get_or_insert_with(|| discovery_shutdown_error(&err));
+                    index += 1;
+                }
+            }
+        }
+        match first_error {
+            Some(err) => Err(err),
+            None => Ok(()),
+        }
+    }
+}
+
+fn discovery_shutdown_error(err: &McpClientError) -> io::Error {
+    io::Error::other(bounded_diagnostic(format!(
+        "MCP discovery shutdown unconfirmed; client retained for cleanup retry: {err}"
+    )))
+}
+
+/// Restore preparation must fail if discovery cannot confirm process shutdown.
+/// The caller retains `cleanup` before this future is polled, including during
+/// initialization. Ordinary discovery failures remain isolated contributions
+/// only after their client has been reaped.
+pub async fn discover_stdio_tool_feature_with_cleanup(
+    config: &McpConfig,
+    workspace_root: &Path,
+    cleanup: &McpDiscoveryCleanup,
+) -> Result<Option<McpStdioToolFeature>, io::Error> {
+    if config.stdio_servers.is_empty() {
+        return Ok(None);
+    }
+    let mut feature = McpStdioToolFeature::new();
+    for server in &config.stdio_servers {
+        match resolve_stdio_server(server, workspace_root, None) {
+            Ok(spec) => {
+                feature.add_contribution(
+                    discover_server_tools_with_cleanup(spec, cleanup, McpStdioLimits::default())
+                        .await?,
+                );
+            }
+            Err(err) => {
+                feature.add_diagnostic(FeatureDiagnostic::error(bounded_diagnostic(format!(
+                    "failed to resolve MCP stdio server `{}`: {err}",
+                    server.name
+                ))))
+            }
+        }
+    }
+    Ok(Some(feature))
+}
+
 /// Discover enabled MCP stdio server tools and return a single feature module
 /// containing startup contributions for normal ToolRegistry installation.
+/// Callers needing retryable startup cleanup must use the cleanup-aware API.
 pub async fn discover_stdio_tool_feature(
     config: &McpConfig,
     workspace_root: &Path,
@@ -81,21 +163,66 @@ pub async fn discover_stdio_tool_feature(
 }
 
 async fn discover_server_tools(spec: McpStdioServerSpec) -> ProtocolProviderContribution {
+    let declaration = provider_declaration(&spec.name, None);
+    let cleanup = McpDiscoveryCleanup::new();
+    match discover_server_tools_with_cleanup(spec, &cleanup, McpStdioLimits::default()).await {
+        Ok(contribution) => contribution,
+        Err(err) => {
+            ProtocolProviderContribution::failed(declaration, bounded_diagnostic(err.to_string()))
+        }
+    }
+}
+
+async fn discover_server_tools_with_cleanup(
+    spec: McpStdioServerSpec,
+    cleanup: &McpDiscoveryCleanup,
+    limits: McpStdioLimits,
+) -> io::Result<ProtocolProviderContribution> {
+    let declaration = provider_declaration(&spec.name, None);
+    let mut clients = cleanup.clients.lock().await;
+    if !clients.is_empty() {
+        return Err(io::Error::other(
+            "MCP discovery cleanup is pending; shut down the retained owner before retrying discovery",
+        ));
+    }
+    let client = match McpStdioClient::spawn_uninitialized(spec.clone(), limits) {
+        Ok(client) => client,
+        Err(err) => {
+            // The synchronous spawn API only errors before a process exists;
+            // pipe/handshake failures retain a client and use the shutdown path.
+            return Ok(ProtocolProviderContribution::failed(
+                declaration,
+                bounded_diagnostic(err.to_string()),
+            ));
+        }
+    };
+    // Spawn is synchronous: the owner holds the client before the first await
+    // that can leave a running process on initialization failure/cancellation.
+    clients.push(client);
+    let client = clients.last_mut().expect("retained discovery client");
+    let contribution = match client.initialize_with_timeout().await {
+        Ok(()) => discover_connected_server_tools(spec, client).await,
+        Err(err) => {
+            ProtocolProviderContribution::failed(declaration, bounded_diagnostic(err.to_string()))
+        }
+    };
+    client
+        .shutdown()
+        .await
+        .map_err(|err| discovery_shutdown_error(&err))?;
+    drop(clients.pop());
+    Ok(contribution)
+}
+
+async fn discover_connected_server_tools(
+    spec: McpStdioServerSpec,
+    client: &mut McpStdioClient,
+) -> ProtocolProviderContribution {
     let declaration = provider_declaration(&spec.name, None);
     let mut contribution = ProtocolProviderContribution::ready(declaration.clone());
     let server_namespace = sanitize_segment(&spec.name);
     let execution_spec = spec.clone();
 
-    let mut client = match McpStdioClient::connect(spec, McpStdioLimits::default()).await {
-        Ok(client) => client,
-        Err(err) => {
-            return ProtocolProviderContribution::failed(
-                declaration,
-                bounded_diagnostic(err.to_string()),
-            );
-        }
-    };
-
     let (server_version, has_tools, has_resources, has_prompts) = client
         .initialize_result()
         .map(|result| {
@@ -133,18 +260,10 @@ async fn discover_server_tools(spec: McpStdioServerSpec) -> ProtocolProviderCont
         {
             Ok(list) => list,
             Err(err) => {
-                let mut failed = ProtocolProviderContribution::failed(
+                return ProtocolProviderContribution::failed(
                     declaration,
                     bounded_diagnostic(err.to_string()),
                 );
-                if let Err(shutdown_err) = client.shutdown().await {
-                    failed = failed.with_diagnostic(FeatureDiagnostic::warning(
-                        bounded_diagnostic(format!(
-                            "MCP server shutdown after discovery failure failed: {shutdown_err}"
-                        )),
-                    ));
-                }
-                return failed;
             }
         };
         let list_changes = client.snapshot_list_changes().await;
@@ -166,20 +285,12 @@ async fn discover_server_tools(spec: McpStdioServerSpec) -> ProtocolProviderCont
             {
                 Ok(list) => list,
                 Err(err) => {
-                    let mut failed = ProtocolProviderContribution::failed(
+                    return ProtocolProviderContribution::failed(
                         declaration,
                         bounded_diagnostic(format!(
                             "MCP server sent notifications/tools/list_changed during tool discovery and refresh failed: {err}"
                         )),
                     );
-                    if let Err(shutdown_err) = client.shutdown().await {
-                        failed = failed.with_diagnostic(FeatureDiagnostic::warning(
-                            bounded_diagnostic(format!(
-                                "MCP server shutdown after discovery refresh failure failed: {shutdown_err}"
-                            )),
-                        ));
-                    }
-                    return failed;
                 }
             };
             let refresh_changes = client.snapshot_list_changes().await;
@@ -197,13 +308,6 @@ async fn discover_server_tools(spec: McpStdioServerSpec) -> ProtocolProviderCont
     } else {
         None
     };
-    let shutdown_result = client.shutdown().await;
-    if let Err(err) = shutdown_result {
-        contribution =
-            contribution.with_diagnostic(FeatureDiagnostic::warning(bounded_diagnostic(format!(
-                "MCP server shutdown after capability discovery failed: {err}"
-            ))));
-    }
 
     if let Some(list) = list {
         contribution = normalize_listed_tools(
@@ -1819,6 +1923,140 @@ read exit_notification || true
         McpStdioServerSpec::new("shell-mock", "/bin/sh").args(["-c".to_string(), script])
     }
 
+    #[cfg(unix)]
+    fn discovery_config(spec: McpStdioServerSpec) -> McpConfig {
+        McpConfig {
+            stdio_servers: vec![manifest::McpStdioServerConfig {
+                name: spec.name,
+                command: spec.command,
+                args: spec.args,
+                cwd: None,
+                env: Default::default(),
+            }],
+        }
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn cancelled_discovery_retains_initializing_client_in_cleanup_clone() {
+        use std::future::{Future, poll_fn};
+        use std::task::Poll;
+        use std::time::Duration;
+
+        let config = discovery_config(
+            McpStdioServerSpec::new("initializing", "/bin/sh")
+                .args(["-c", "while read line; do :; done"]),
+        );
+        let cleanup = McpDiscoveryCleanup::new();
+        let retained = cleanup.clone();
+        let mut discovery = Box::pin(discover_stdio_tool_feature_with_cleanup(
+            &config,
+            Path::new("."),
+            &cleanup,
+        ));
+        poll_fn(|cx| {
+            assert!(discovery.as_mut().poll(cx).is_pending());
+            Poll::Ready(())
+        })
+        .await;
+        drop(discovery);
+        drop(cleanup);
+        assert_eq!(retained.clients.lock().await.len(), 1);
+        let retry_error =
+            discover_stdio_tool_feature_with_cleanup(&config, Path::new("."), &retained)
+                .await
+                .err()
+                .expect("discovery must not replace an unreaped client");
+        assert!(retry_error.to_string().contains("cleanup is pending"));
+        assert_eq!(retained.clients.lock().await.len(), 1);
+        tokio::time::timeout(Duration::from_secs(3), retained.shutdown())
+            .await
+            .expect("cleanup bounded")
+            .expect("initializing process reaped");
+        assert!(retained.clients.lock().await.is_empty());
+        retained
+            .shutdown()
+            .await
+            .expect("confirmed cleanup repeatable");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn cancelled_owner_shutdown_keeps_client_for_retry() {
+        use std::future::{Future, poll_fn};
+        use std::task::Poll;
+        use std::time::Duration;
+
+        let cleanup = McpDiscoveryCleanup::new();
+        cleanup.clients.lock().await.push(
+            McpStdioClient::spawn_uninitialized(
+                McpStdioServerSpec::new("shutdown-pending", "/bin/sh")
+                    .args(["-c", "while :; do :; done"]),
+                McpStdioLimits {
+                    shutdown_timeout: Duration::from_millis(50),
+                    kill_timeout: Duration::from_millis(100),
+                    ..Default::default()
+                },
+            )
+            .expect("local process spawned"),
+        );
+        let mut shutdown = Box::pin(cleanup.shutdown());
+        poll_fn(|cx| {
+            assert!(shutdown.as_mut().poll(cx).is_pending());
+            Poll::Ready(())
+        })
+        .await;
+        drop(shutdown);
+        assert_eq!(cleanup.clients.lock().await.len(), 1);
+        tokio::time::timeout(Duration::from_secs(2), cleanup.shutdown())
+            .await
+            .expect("retry bounded")
+            .expect("retained client shutdown");
+        assert!(cleanup.clients.lock().await.is_empty());
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn cleanup_aware_discovery_reaps_success_and_initialize_failure() {
+        let cleanup = McpDiscoveryCleanup::new();
+        let config = discovery_config(shell_capability_server(r#"{"resources":{}}"#));
+        let feature = discover_stdio_tool_feature_with_cleanup(&config, Path::new("."), &cleanup)
+            .await
+            .expect("discovery cleanup confirmed")
+            .expect("configured feature");
+        assert!(
+            feature.contributions[0]
+                .tools
+                .iter()
+                .any(|tool| { tool.name() == "Mcp_shell_mock_resources_read" })
+        );
+        assert!(cleanup.clients.lock().await.is_empty());
+
+        let spec = McpStdioServerSpec::new("init-rejected", "/bin/sh").args([
+            "-c",
+            r#"read init || exit 1
+printf '%s\n' '{"jsonrpc":"2.0","id":1,"error":{"code":-32000,"message":"initialization rejected"}}'
+while read line; do :; done
+"#,
+        ]);
+        let config = discovery_config(spec);
+        let feature = discover_stdio_tool_feature_with_cleanup(&config, Path::new("."), &cleanup)
+            .await
+            .expect("failure isolated only after cleanup")
+            .expect("failed contribution preserved");
+        assert!(feature.contributions[0].tools.is_empty());
+        assert!(
+            feature.contributions[0]
+                .diagnostics
+                .iter()
+                .any(|diagnostic| {
+                    diagnostic.severity == FeatureDiagnosticSeverity::Error
+                        && diagnostic.message.contains("initialization rejected")
+                })
+        );
+        assert!(cleanup.clients.lock().await.is_empty());
+    }
+
     #[test]
     fn mcp_resource_prompt_operation_tools_are_namespaced_and_untrusted() {
         let declaration = provider_declaration("demo", Some("1.0.0"));

crates/worker/src/lib.rs

@@ -19,6 +19,9 @@ pub mod model_client;
 mod paste_artifact_tool;
 pub mod profile_manifest;
 pub mod prompt;
+#[cfg(test)]
+#[path = "../../../tests/support/restore_environment.rs"]
+mod restore_test_environment;
 pub mod runtime;
 pub mod segment_log_sink;
 mod session_capture;
@@ -70,9 +73,10 @@ pub use session_history::{
 pub use shared_state::WorkerSharedState;
 pub use worker::{
     LocalWorkingDirectory, SubjektivSessionAttributionLifecycle, Worker, WorkerError,
-    WorkerFilesystemAuthority, WorkerRunResult, WorkerWorkspaceContext, WorkspaceBinaryRequest,
-    WorkspaceBinaryResponse, WorkspaceClient, WorkspaceClientError, WorkspaceId, WorkspaceIdError,
-    WorkspacePromptCatalogResolution, WorkspaceRequest, WorkspaceRequestMethod, WorkspaceResponse,
-    WorkspaceServerOperation, WorkspaceWorkerDiscoveryRequest, apply_worker_manifest,
-    marker_workspace_client, unavailable_workspace_client,
+    WorkerFilesystemAuthority, WorkerRunResult, WorkerStartupFeatureCleanup,
+    WorkerWorkspaceContext, WorkspaceBinaryRequest, WorkspaceBinaryResponse, WorkspaceClient,
+    WorkspaceClientError, WorkspaceId, WorkspaceIdError, WorkspacePromptCatalogResolution,
+    WorkspaceRequest, WorkspaceRequestMethod, WorkspaceResponse, WorkspaceServerOperation,
+    WorkspaceWorkerDiscoveryRequest, apply_worker_manifest, marker_workspace_client,
+    unavailable_workspace_client,
 };

crates/worker/src/runtime/worker_allocation.rs

@@ -21,8 +21,9 @@ pub use error::ScopeLockError;
 #[cfg(test)]
 pub(crate) use lifecycle::install_top_level_at_for_test;
 pub use lifecycle::{
-    ScopeAllocationGuard, SegmentActivationCommitError, SegmentActivationGuard, SegmentLockInfo,
-    adopt_allocation, install_top_level, install_top_level_with_deny, lookup_segment,
+    ScopeAllocationCleanup, ScopeAllocationGuard, SegmentActivationCommitError,
+    SegmentActivationGuard, SegmentLockInfo, adopt_allocation, install_top_level,
+    install_top_level_with_deny, lookup_segment,
 };
 pub use mutate::{
     delegate_scope, reclaim_delegated_scope, reclaim_stale, reclaim_stale_with, register_worker,

crates/worker/src/runtime/worker_allocation/error.rs

@@ -11,6 +11,14 @@ use session_store::SegmentId;
 pub enum ScopeLockError {
     #[error("I/O error on workers.json: {0}")]
     Io(#[from] io::Error),
+    #[error("worker `{worker_name}` allocation still has a live guard")]
+    AllocationGuardAttached { worker_name: String },
+    #[error("allocation acquisition was published but its save failed: {source}")]
+    AcquisitionSaveFailed {
+        #[source]
+        source: io::Error,
+        cleanup: super::lifecycle::ScopeAllocationCleanup,
+    },
     #[error("worker `{0}` is already allocated")]
     DuplicateWorkerName(String),
     #[error("requested scope `{}` conflicts with worker allocation `{competitor}` rule `{}`", .rule.target.display(), .competitor_rule.target.display())]
@@ -46,3 +54,15 @@ pub enum ScopeLockError {
         actual: Option<SegmentId>,
     },
 }
+
+impl ScopeLockError {
+    /// Retain the release authority for an owned acquisition whose table rename
+    /// succeeded before save returned an error. Refusals and unpublished saves
+    /// do not confer ownership of any preexisting allocation.
+    pub fn allocation_cleanup_handle(&self) -> Option<super::lifecycle::ScopeAllocationCleanup> {
+        match self {
+            Self::AcquisitionSaveFailed { cleanup, .. } => Some(cleanup.clone()),
+            _ => None,
+        }
+    }
+}

crates/worker/src/runtime/worker_allocation/lifecycle.rs

@@ -3,6 +3,7 @@
 //! up on drop.
 
 use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex};
 
 use manifest::ScopeRule;
 use session_store::SegmentId;
@@ -11,13 +12,82 @@ use super::error::ScopeLockError;
 use super::mutate::release_worker;
 use super::table::{LockFileGuard, default_allocation_path};
 
-/// Owned allocation: on drop, opens the lock file and releases this
-/// Worker's entry. The guard keeps only the name + lock-file path; it
-/// does not hold the `flock` for the Worker's lifetime.
+/// Owned allocation shared with any retained cleanup handles. Only the final
+/// owner's drop attempts best-effort release; no `flock` is held for its lifetime.
 #[derive(Debug)]
 pub struct ScopeAllocationGuard {
+    owner: Arc<ScopeAllocationOwner>,
+}
+
+/// Retains an allocation's release authority even after its Worker is consumed
+/// or dropped. A successful controller transfer should drop this handle without
+/// explicitly releasing the live Worker's allocation.
+#[derive(Debug, Clone)]
+pub struct ScopeAllocationCleanup {
+    owner: Arc<ScopeAllocationOwner>,
+}
+
+#[derive(Debug)]
+struct ScopeAllocationOwner {
     worker_name: String,
     lock_path: PathBuf,
+    state: Mutex<ScopeAllocationOwnerState>,
+}
+
+#[derive(Debug)]
+struct ScopeAllocationOwnerState {
+    guard_attached: bool,
+    released: bool,
+}
+
+impl ScopeAllocationOwner {
+    fn release(&self) -> Result<(), ScopeLockError> {
+        let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+        if state.guard_attached {
+            return Err(ScopeLockError::AllocationGuardAttached {
+                worker_name: self.worker_name.clone(),
+            });
+        }
+        if state.released {
+            return Ok(());
+        }
+        let mut guard = LockFileGuard::open(&self.lock_path)?;
+        if guard.data().find(&self.worker_name).is_some() {
+            release_worker(&mut guard, &self.worker_name)?;
+        } else {
+            // A previous save may have removed the entry but failed after rename.
+            // Complete the durable save before reporting release as confirmed.
+            guard.save()?;
+        }
+        state.released = true;
+        Ok(())
+    }
+}
+
+impl Drop for ScopeAllocationOwner {
+    fn drop(&mut self) {
+        let _ = self.release();
+    }
+}
+
+impl Drop for ScopeAllocationGuard {
+    fn drop(&mut self) {
+        self.owner
+            .state
+            .lock()
+            .unwrap_or_else(|error| error.into_inner())
+            .guard_attached = false;
+    }
+}
+
+impl ScopeAllocationCleanup {
+    /// Release only after the factory has returned an error and no live
+    /// Controller can still use this allocation. An attached guard rejects
+    /// release. Success is idempotent across all clones; failure retains the
+    /// authority for an explicit retry.
+    pub fn release(&self) -> Result<(), ScopeLockError> {
+        self.owner.release()
+    }
 }
 
 /// Exclusive authority to activate a replacement Segment for one allocated Worker.
@@ -56,12 +126,58 @@ impl SegmentActivationCommitError {
 }
 
 impl ScopeAllocationGuard {
+    fn new(worker_name: String, lock_path: PathBuf) -> Self {
+        Self {
+            owner: Arc::new(ScopeAllocationOwner {
+                worker_name,
+                lock_path,
+                state: Mutex::new(ScopeAllocationOwnerState {
+                    guard_attached: true,
+                    released: false,
+                }),
+            }),
+        }
+    }
+
+    fn commit_acquisition(self, table: &mut LockFileGuard) -> Result<Self, ScopeLockError> {
+        let mut published = false;
+        match table.save_with_publication(&mut published) {
+            Ok(()) => Ok(self),
+            Err(source) => {
+                let cleanup = if published {
+                    Some(self.cleanup_handle())
+                } else {
+                    // A staged entry is not ours until rename. Disarm before
+                    // detaching so Drop cannot remove a preexisting allocation
+                    // (or reopen the table lock still held by this caller).
+                    self.owner
+                        .state
+                        .lock()
+                        .unwrap_or_else(|error| error.into_inner())
+                        .released = true;
+                    None
+                };
+                drop(self);
+                match cleanup {
+                    Some(cleanup) => Err(ScopeLockError::AcquisitionSaveFailed { source, cleanup }),
+                    None => Err(ScopeLockError::Io(source)),
+                }
+            }
+        }
+    }
+
     pub fn worker_name(&self) -> &str {
-        &self.worker_name
+        &self.owner.worker_name
     }
 
     pub fn lock_path(&self) -> &Path {
-        &self.lock_path
+        &self.owner.lock_path
+    }
+
+    pub fn cleanup_handle(&self) -> ScopeAllocationCleanup {
+        ScopeAllocationCleanup {
+            owner: Arc::clone(&self.owner),
+        }
     }
 
     /// Lock the machine-wide table and validate a Segment replacement before
@@ -71,24 +187,25 @@ impl ScopeAllocationGuard {
         expected: SegmentId,
         replacement: SegmentId,
     ) -> Result<SegmentActivationGuard, ScopeLockError> {
-        let guard = LockFileGuard::open(&self.lock_path)?;
+        let guard = LockFileGuard::open(self.lock_path())?;
         let actual = guard
             .data()
             .allocations
             .iter()
-            .find(|allocation| allocation.worker_name == self.worker_name)
+            .find(|allocation| allocation.worker_name == self.worker_name())
             .and_then(|allocation| allocation.segment_id);
 
         if actual != Some(expected) {
             return Err(ScopeLockError::SegmentChanged {
-                worker_name: self.worker_name.clone(),
+                worker_name: self.worker_name().to_owned(),
                 expected,
                 actual,
             });
         }
 
         if let Some(existing) = guard.data().allocations.iter().find(|allocation| {
-            allocation.segment_id == Some(replacement) && allocation.worker_name != self.worker_name
+            allocation.segment_id == Some(replacement)
+                && allocation.worker_name != self.worker_name()
         }) {
             return Err(ScopeLockError::SegmentConflict {
                 segment_id: replacement,
@@ -99,7 +216,7 @@ impl ScopeAllocationGuard {
 
         Ok(SegmentActivationGuard {
             guard,
-            worker_name: self.worker_name.clone(),
+            worker_name: self.worker_name().to_owned(),
             expected,
             replacement,
         })
@@ -145,14 +262,6 @@ impl SegmentActivationGuard {
     }
 }
 
-impl Drop for ScopeAllocationGuard {
-    fn drop(&mut self) {
-        if let Ok(mut guard) = LockFileGuard::open(&self.lock_path) {
-            let _ = release_worker(&mut guard, &self.worker_name);
-        }
-    }
-}
-
 /// Open the default lock file, register a top-level Worker, and return a
 /// guard that will release the allocation on drop.
 pub fn install_top_level(
@@ -185,7 +294,7 @@ pub fn install_top_level_with_deny(
 ) -> Result<ScopeAllocationGuard, ScopeLockError> {
     let lock_path = default_allocation_path()?;
     let mut guard = LockFileGuard::open(&lock_path)?;
-    super::mutate::register_worker_with_deny(
+    super::mutate::prepare_worker_with_deny(
         &mut guard,
         worker_name.clone(),
         pid,
@@ -194,10 +303,7 @@ pub fn install_top_level_with_deny(
         scope_deny,
         segment_id,
     )?;
-    Ok(ScopeAllocationGuard {
-        worker_name,
-        lock_path,
-    })
+    ScopeAllocationGuard::new(worker_name, lock_path).commit_acquisition(&mut guard)
 }
 
 #[cfg(test)]
@@ -210,7 +316,7 @@ pub(crate) fn install_top_level_at_for_test(
     segment_id: SegmentId,
 ) -> Result<ScopeAllocationGuard, ScopeLockError> {
     let mut guard = LockFileGuard::open(&lock_path)?;
-    super::mutate::register_worker_with_deny(
+    super::mutate::prepare_worker_with_deny(
         &mut guard,
         worker_name.clone(),
         pid,
@@ -219,10 +325,7 @@ pub(crate) fn install_top_level_at_for_test(
         Vec::new(),
         segment_id,
     )?;
-    Ok(ScopeAllocationGuard {
-        worker_name,
-        lock_path,
-    })
+    ScopeAllocationGuard::new(worker_name, lock_path).commit_acquisition(&mut guard)
 }
 
 /// Take ownership of an existing allocation that was pre-registered by
@@ -240,6 +343,15 @@ pub fn adopt_allocation(
     segment_id: SegmentId,
 ) -> Result<ScopeAllocationGuard, ScopeLockError> {
     let lock_path = default_allocation_path()?;
+    adopt_allocation_at(lock_path, worker_name, new_pid, segment_id)
+}
+
+fn adopt_allocation_at(
+    lock_path: PathBuf,
+    worker_name: String,
+    new_pid: u32,
+    segment_id: SegmentId,
+) -> Result<ScopeAllocationGuard, ScopeLockError> {
     let mut guard = LockFileGuard::open(&lock_path)?;
     let alloc = guard
         .data_mut()
@@ -247,11 +359,7 @@ pub fn adopt_allocation(
         .ok_or_else(|| ScopeLockError::UnknownWorker(worker_name.clone()))?;
     alloc.pid = new_pid;
     alloc.segment_id = Some(segment_id);
-    guard.save()?;
-    Ok(ScopeAllocationGuard {
-        worker_name,
-        lock_path,
-    })
+    ScopeAllocationGuard::new(worker_name, lock_path).commit_acquisition(&mut guard)
 }
 
 /// Information about a Worker that currently holds an allocation for a
@@ -329,6 +437,388 @@ mod tests {
         }
     }
 
+    #[test]
+    fn retained_cleanup_handle_defers_release_until_final_owner_drop() {
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let guard = install_top_level_at_for_test(
+            lock_path.clone(),
+            "retained".into(),
+            std::process::id(),
+            sock("retained"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap();
+        let cleanup = guard.cleanup_handle();
+        let retained = cleanup.clone();
+
+        drop(guard);
+        drop(cleanup);
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert!(table.data().find("retained").is_some());
+        }
+        drop(retained);
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("retained").is_none());
+    }
+
+    #[test]
+    fn dropping_cleanup_handle_after_transfer_preserves_live_guard() {
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let segment = sid();
+        let guard = install_top_level_at_for_test(
+            lock_path.clone(),
+            "transferred".into(),
+            std::process::id(),
+            sock("transferred"),
+            Vec::new(),
+            segment,
+        )
+        .unwrap();
+
+        drop(guard.cleanup_handle());
+        // A successful controller transfer leaves the guard usable, including
+        // the existing activation protocol, without retaining a cleanup handle.
+        let replacement = sid();
+        guard
+            .begin_segment_activation(segment, replacement)
+            .unwrap()
+            .commit()
+            .unwrap();
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert_eq!(
+                table.data().find("transferred").unwrap().segment_id,
+                Some(replacement)
+            );
+        }
+        drop(guard);
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("transferred").is_none());
+    }
+
+    #[test]
+    fn cleanup_rejects_live_guard_then_releases_only_detached_allocation() {
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let guard = install_top_level_at_for_test(
+            lock_path.clone(),
+            "live".into(),
+            std::process::id(),
+            sock("live"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap();
+        let _competitor = install_top_level_at_for_test(
+            lock_path.clone(),
+            "competitor".into(),
+            std::process::id(),
+            sock("competitor"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap();
+        let cleanup = guard.cleanup_handle();
+        assert!(matches!(
+            cleanup.release(),
+            Err(ScopeLockError::AllocationGuardAttached { worker_name }) if worker_name == "live"
+        ));
+        drop(cleanup.clone());
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert!(table.data().find("live").is_some());
+            assert!(table.data().find("competitor").is_some());
+        }
+
+        drop(guard);
+        cleanup.release().unwrap();
+        cleanup.clone().release().unwrap();
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("live").is_none());
+        assert!(table.data().find("competitor").is_some());
+    }
+
+    #[test]
+    fn published_install_failure_retains_cleanup_across_error_drop_and_retry() {
+        use super::super::table::{SaveFaultPoint, fail_next_save_at_for_test};
+
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let _competitor = install_top_level_at_for_test(
+            lock_path.clone(),
+            "competitor".into(),
+            std::process::id(),
+            sock("competitor"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap();
+        let segment = sid();
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::AfterRename);
+        let error = install_top_level_at_for_test(
+            lock_path.clone(),
+            "failed".into(),
+            std::process::id(),
+            sock("failed"),
+            Vec::new(),
+            segment,
+        )
+        .unwrap_err();
+        assert!(matches!(
+            &error,
+            ScopeLockError::AcquisitionSaveFailed { .. }
+        ));
+        let cleanup = error.allocation_cleanup_handle().unwrap();
+        let retry = error.allocation_cleanup_handle().unwrap();
+        drop(error);
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert_eq!(
+                table.data().find("failed").unwrap().segment_id,
+                Some(segment)
+            );
+            assert!(table.data().find("competitor").is_some());
+        }
+
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::BeforeRename);
+        assert!(matches!(cleanup.release(), Err(ScopeLockError::Io(_))));
+        drop(cleanup);
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert!(table.data().find("failed").is_some());
+        }
+        retry.release().unwrap();
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("failed").is_none());
+        assert!(table.data().find("competitor").is_some());
+    }
+
+    #[test]
+    fn refused_install_has_no_cleanup_authority_over_existing_allocation() {
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let segment = sid();
+        let _competitor = install_top_level_at_for_test(
+            lock_path.clone(),
+            "competitor".into(),
+            std::process::id(),
+            sock("competitor"),
+            vec![write_rule("/shared", true)],
+            segment,
+        )
+        .unwrap();
+        let cases = [
+            ("competitor", sid(), Vec::new()),
+            ("other", segment, Vec::new()),
+            ("other", sid(), vec![write_rule("/shared", true)]),
+        ];
+        for (name, requested_segment, scope) in cases {
+            let error = install_top_level_at_for_test(
+                lock_path.clone(),
+                name.into(),
+                std::process::id(),
+                sock("refused"),
+                scope,
+                requested_segment,
+            )
+            .unwrap_err();
+            assert!(matches!(
+                &error,
+                ScopeLockError::DuplicateWorkerName(_)
+                    | ScopeLockError::SegmentConflict { .. }
+                    | ScopeLockError::WriteConflict { .. }
+            ));
+            assert!(error.allocation_cleanup_handle().is_none());
+            drop(error);
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert_eq!(table.data().allocations.len(), 1);
+            assert_eq!(
+                table.data().find("competitor").unwrap().socket,
+                sock("competitor")
+            );
+            assert_eq!(
+                table.data().find("competitor").unwrap().segment_id,
+                Some(segment)
+            );
+        }
+    }
+
+    #[test]
+    fn unpublished_install_failure_has_no_cleanup_authority() {
+        use super::super::table::{SaveFaultPoint, fail_next_save_at_for_test};
+
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let _competitor = install_top_level_at_for_test(
+            lock_path.clone(),
+            "competitor".into(),
+            std::process::id(),
+            sock("competitor"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap();
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::BeforeRename);
+        let error = install_top_level_at_for_test(
+            lock_path.clone(),
+            "failed".into(),
+            std::process::id(),
+            sock("failed"),
+            Vec::new(),
+            sid(),
+        )
+        .unwrap_err();
+        assert!(matches!(&error, ScopeLockError::Io(_)));
+        assert!(error.allocation_cleanup_handle().is_none());
+        drop(error);
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("failed").is_none());
+        assert!(table.data().find("competitor").is_some());
+    }
+
+    #[test]
+    fn published_adoption_failure_retains_cleanup_for_retry() {
+        use super::super::table::{SaveFaultPoint, fail_next_save_at_for_test};
+
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        {
+            let mut table = LockFileGuard::open(&lock_path).unwrap();
+            delegate_placeholder(&mut table, "child", std::process::id());
+            delegate_placeholder(&mut table, "competitor", std::process::id());
+        }
+        let segment = sid();
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::AfterRename);
+        let error = adopt_allocation_at(
+            lock_path.clone(),
+            "child".into(),
+            std::process::id(),
+            segment,
+        )
+        .unwrap_err();
+        let cleanup = error.allocation_cleanup_handle().unwrap();
+        drop(error);
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert_eq!(
+                table.data().find("child").unwrap().segment_id,
+                Some(segment)
+            );
+        }
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::AfterRename);
+        assert!(matches!(cleanup.release(), Err(ScopeLockError::Io(_))));
+        {
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert!(table.data().find("child").is_none());
+            assert!(table.data().find("competitor").is_some());
+        }
+        cleanup.release().unwrap();
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        assert!(table.data().find("child").is_none());
+        assert!(table.data().find("competitor").is_some());
+    }
+
+    #[test]
+    fn unpublished_adoption_failure_preserves_placeholder_without_cleanup_authority() {
+        use super::super::table::{SaveFaultPoint, fail_next_save_at_for_test};
+
+        let dir = TempDir::new().unwrap();
+        let lock_path = dir.path().join("workers.json");
+        let placeholder_pid = std::process::id();
+        {
+            let mut table = LockFileGuard::open(&lock_path).unwrap();
+            delegate_placeholder(&mut table, "child", placeholder_pid);
+        }
+        fail_next_save_at_for_test(&lock_path, SaveFaultPoint::BeforeRename);
+        let error = adopt_allocation_at(
+            lock_path.clone(),
+            "child".into(),
+            placeholder_pid.wrapping_add(1),
+            sid(),
+        )
+        .unwrap_err();
+        assert!(matches!(&error, ScopeLockError::Io(_)));
+        assert!(error.allocation_cleanup_handle().is_none());
+        drop(error);
+        let table = LockFileGuard::open(&lock_path).unwrap();
+        let allocation = table.data().find("child").unwrap();
+        assert_eq!(allocation.pid, placeholder_pid);
+        assert_eq!(allocation.segment_id, None);
+    }
+
+    #[test]
+    fn failed_cleanup_release_retains_authority_for_retry() {
+        use super::super::table::{SaveFaultPoint, fail_next_save_at_for_test};
+
+        for point in [SaveFaultPoint::BeforeRename, SaveFaultPoint::AfterRename] {
+            let dir = TempDir::new().unwrap();
+            let lock_path = dir.path().join("workers.json");
+            let guard = install_top_level_at_for_test(
+                lock_path.clone(),
+                "failed".into(),
+                std::process::id(),
+                sock("failed"),
+                vec![read_rule("/shared", true)],
+                sid(),
+            )
+            .unwrap();
+            let _other = install_top_level_at_for_test(
+                lock_path.clone(),
+                "other".into(),
+                std::process::id(),
+                sock("other"),
+                vec![read_rule("/shared", true)],
+                sid(),
+            )
+            .unwrap();
+            let cleanup = guard.cleanup_handle();
+            let retry = cleanup.clone();
+            drop(guard);
+            fail_next_save_at_for_test(&lock_path, point);
+
+            assert!(matches!(cleanup.release(), Err(ScopeLockError::Io(_))));
+            drop(cleanup);
+            {
+                let table = LockFileGuard::open(&lock_path).unwrap();
+                assert_eq!(
+                    table.data().find("failed").is_some(),
+                    point == SaveFaultPoint::BeforeRename
+                );
+                assert!(table.data().find("other").is_some());
+            }
+            retry.release().unwrap();
+            {
+                let table = LockFileGuard::open(&lock_path).unwrap();
+                assert!(table.data().find("failed").is_none());
+                assert!(table.data().find("other").is_some());
+            }
+
+            // Success disarms every clone and final Drop, even if the name is
+            // subsequently allocated to a different Worker.
+            let _replacement = install_top_level_at_for_test(
+                lock_path.clone(),
+                "failed".into(),
+                std::process::id(),
+                sock("replacement"),
+                vec![read_rule("/shared", true)],
+                sid(),
+            )
+            .unwrap();
+            retry.clone().release().unwrap();
+            drop(retry);
+            let table = LockFileGuard::open(&lock_path).unwrap();
+            assert_eq!(
+                table.data().find("failed").unwrap().socket,
+                sock("replacement")
+            );
+            assert!(table.data().find("other").is_some());
+        }
+    }
+
     #[test]
     fn adopt_allocation_rewrites_pid_and_releases_on_drop() {
         let dir = TempDir::new().unwrap();
@@ -358,6 +848,7 @@ mod tests {
         let dir = TempDir::new().unwrap();
         let _sandbox = RuntimeDirSandbox::new(dir.path());
         let err = adopt_allocation("ghost".into(), 42, sid()).unwrap_err();
+        assert!(err.allocation_cleanup_handle().is_none());
         assert!(matches!(err, ScopeLockError::UnknownWorker(ref n) if n == "ghost"));
     }
 

crates/worker/src/runtime/worker_allocation/mutate.rs

@@ -57,6 +57,30 @@ pub fn register_worker_with_deny(
     scope_allow: Vec<ScopeRule>,
     scope_deny: Vec<ScopeRule>,
     segment_id: SegmentId,
+) -> Result<(), ScopeLockError> {
+    prepare_worker_with_deny(
+        guard,
+        worker_name,
+        pid,
+        socket,
+        scope_allow,
+        scope_deny,
+        segment_id,
+    )?;
+    guard.save()?;
+    Ok(())
+}
+
+/// Admit and stage a registration without publishing it. Owned installation
+/// constructs its cleanup authority after admission and before committing.
+pub(super) fn prepare_worker_with_deny(
+    guard: &mut LockFileGuard,
+    worker_name: String,
+    pid: u32,
+    socket: PathBuf,
+    scope_allow: Vec<ScopeRule>,
+    scope_deny: Vec<ScopeRule>,
+    segment_id: SegmentId,
 ) -> Result<(), ScopeLockError> {
     reclaim_stale(guard);
     if guard.data().find(&worker_name).is_some() {
@@ -101,7 +125,6 @@ pub fn register_worker_with_deny(
         delegated_from: None,
         segment_id: Some(segment_id),
     });
-    guard.save()?;
     Ok(())
 }
 

crates/worker/src/runtime/worker_allocation/table.rs

@@ -246,6 +246,14 @@ impl LockFileGuard {
     /// Persist with atomic replacement while the separate allocation lock stays
     /// held. A crash exposes either the old complete table or the new one.
     pub fn save(&mut self) -> io::Result<()> {
+        let mut published = false;
+        self.save_with_publication(&mut published)
+    }
+
+    /// Report this save's rename while the caller still holds the table lock.
+    /// In-memory mutation alone is not evidence that an acquisition took place.
+    pub(super) fn save_with_publication(&mut self, published: &mut bool) -> io::Result<()> {
+        *published = false;
         let json = serde_json::to_vec_pretty(&self.data).map_err(io::Error::other)?;
         let mut temp_os = self.data_path.as_os_str().to_owned();
         temp_os.push(".tmp");
@@ -265,6 +273,7 @@ impl LockFileGuard {
             ));
         }
         fs::rename(&temp_path, &self.data_path)?;
+        *published = true;
         #[cfg(test)]
         if take_save_fault(&self.data_path, SaveFaultPoint::AfterRename) {
             return Err(io::Error::other(

crates/worker/src/spawn/registry.rs

@@ -340,6 +340,84 @@ pub struct SpawnedWorkerRegistryLoad {
     pub reclaimed_unreachable: bool,
 }
 
+/// Read-only restore preparation using the registry that will be published by the Host.
+/// The Host must retain its lifecycle exclusion from preparation through application;
+/// the metadata reread detects source drift, but is not an atomic store transaction.
+/// Dropping an unapplied plan does not reclaim legacy child state.
+pub(crate) struct PreparedSpawnedWorkerRegistryLoad<St> {
+    pub(crate) registry: Arc<SpawnedWorkerRegistry>,
+    /// Projected outcome, not evidence that reclamation has already happened.
+    pub(crate) reclaimed_unreachable: bool,
+    runtime_dir: Arc<RuntimeDir>,
+    store: St,
+    worker_name: String,
+    parent_scope: Option<SharedScope>,
+    persisted_children: Vec<WorkerSpawnedChild>,
+    valid_records: Vec<SpawnedWorkerRecord>,
+    legacy_projection_exists: bool,
+}
+
+impl<St> PreparedSpawnedWorkerRegistryLoad<St>
+where
+    St: WorkerMetadataStore + Clone + Send + Sync + 'static,
+{
+    /// Apply only after all other fallible tool initialization, before socket publication.
+    /// Effects retain their original order: projection, canonical reclaim history, then
+    /// allocation/scope. An error may follow committed effects; no rollback is attempted.
+    pub(crate) async fn apply(&mut self) -> io::Result<SpawnedWorkerRegistryLoad> {
+        let metadata = self
+            .store
+            .read_by_name(&self.worker_name)
+            .map_err(store_error_to_io)?;
+        let reclaimed: Vec<_> = self
+            .persisted_children
+            .iter()
+            .map(reclaimed_child_from_metadata)
+            .collect();
+        // A prior apply may have committed the canonical reclaim and then failed
+        // in allocation/scope cleanup. Under the host's same lifecycle lock,
+        // confirm its exact retained facts rather than appending them again.
+        let already_reclaimed = !reclaimed.is_empty()
+            && metadata.as_ref().is_some_and(|metadata| {
+                metadata.spawned_children.is_empty()
+                    && metadata.reclaimed_children.ends_with(&reclaimed)
+            });
+        if !already_reclaimed
+            && metadata
+                .as_ref()
+                .map(|metadata| metadata.spawned_children.as_slice())
+                .unwrap_or_default()
+                != self.persisted_children.as_slice()
+        {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidData,
+                "spawned-child metadata changed after registry preparation",
+            ));
+        }
+
+        // Runtime projection is migration input only; the normal Internal registry is never
+        // materialized into spawned_workers.json. Only existence was captured, not its contents.
+        if !self.persisted_children.is_empty() || self.legacy_projection_exists {
+            self.runtime_dir.write_spawned_workers(&[]).await?;
+        }
+        if !reclaimed.is_empty() && !already_reclaimed {
+            self.store
+                .reclaim_spawned_children(&self.worker_name, reclaimed)
+                .map_err(store_error_to_io)?;
+        }
+        for record in &self.valid_records {
+            reclaim_record(&self.worker_name, self.parent_scope.as_ref(), record)?;
+        }
+
+        let reclaimed_unreachable = self.reclaimed_unreachable;
+        self.reclaimed_unreachable = false;
+        Ok(SpawnedWorkerRegistryLoad {
+            registry: self.registry.clone(),
+            reclaimed_unreachable,
+        })
+    }
+}
+
 impl SpawnedWorkerRegistry {
     pub(crate) fn bash_output(&self) -> tools::bash_output::BashOutputView {
         self.bash_output.lock().unwrap().clone()
@@ -426,6 +504,23 @@ impl SpawnedWorkerRegistry {
         worker_name: String,
         parent_scope: Option<SharedScope>,
     ) -> io::Result<SpawnedWorkerRegistryLoad>
+    where
+        St: WorkerMetadataStore + Clone + Send + Sync + 'static,
+    {
+        Self::prepare_from_worker_state_with_reclaim(runtime_dir, store, worker_name, parent_scope)?
+            .apply()
+            .await
+    }
+
+    /// Capture legacy child facts without starting children or changing metadata,
+    /// projection, allocation, or scope. Tool initialization must clone this registry,
+    /// not construct a replacement, and defer reclamation to the owned plan's `apply`.
+    pub(crate) fn prepare_from_worker_state_with_reclaim<St>(
+        runtime_dir: Arc<RuntimeDir>,
+        store: St,
+        worker_name: String,
+        parent_scope: Option<SharedScope>,
+    ) -> io::Result<PreparedSpawnedWorkerRegistryLoad<St>>
     where
         St: WorkerMetadataStore + Clone + Send + Sync + 'static,
     {
@@ -442,54 +537,33 @@ impl SpawnedWorkerRegistry {
                 Ok(record) => {
                     warn!(
                         worker = %record.worker_name,
-                        "reclaiming legacy persisted process SubWorker during Internal session restore"
+                        "captured legacy persisted process SubWorker for reclamation during Internal session restore"
                     );
                     valid_records.push(record);
                 }
                 Err(error) => warn!(
                     error = %error,
                     worker = %child.worker_name,
-                    "clearing corrupt legacy persisted process SubWorker record"
+                    "captured corrupt legacy persisted process SubWorker record for clearing"
                 ),
             }
         }
 
-        // Runtime projection is migration input only; the normal Internal registry is never
-        // materialized into spawned_workers.json.
         let legacy_projection_exists = runtime_dir.path().join("spawned_workers.json").exists();
-        if !persisted_children.is_empty() || legacy_projection_exists {
-            runtime_dir.write_spawned_workers(&[]).await?;
-        }
-        if !persisted_children.is_empty() {
-            let reclaimed = persisted_children
-                .iter()
-                .map(reclaimed_child_from_metadata)
-                .collect();
-            store
-                .reclaim_spawned_children(&worker_name, reclaimed)
-                .map_err(store_error_to_io)?;
-        }
-        for record in &valid_records {
-            reclaim_record(&worker_name, parent_scope.as_ref(), record)?;
-        }
-
-        Ok(SpawnedWorkerRegistryLoad {
-            registry: Arc::new(Self {
-                internal_records: std::sync::Mutex::new(Vec::new()),
-                service_records: std::sync::Mutex::new(Vec::new()),
-                #[cfg(test)]
-                fail_service_stops: std::sync::Mutex::new(HashMap::new()),
-                internal_names: std::sync::Mutex::new(HashSet::new()),
-                internal_shutting_down: AtomicBool::new(false),
-                pending_internal_spawns: AtomicUsize::new(0),
-                pending_internal_notify: Notify::new(),
-                internal_spawn_cleanup_failed: AtomicBool::new(false),
-                parent_scope,
-                bash_output: Mutex::new(tools::bash_output::BashOutputView::new()),
-                parent_protocol: Mutex::new(None),
-                protocol_publish_lock: Arc::new(Mutex::new(())),
-            }),
+        let registry = match &parent_scope {
+            Some(scope) => Self::new_internal(worker_name.clone(), scope.clone()),
+            None => Self::new(runtime_dir.clone()),
+        };
+        Ok(PreparedSpawnedWorkerRegistryLoad {
+            registry,
             reclaimed_unreachable: !persisted_children.is_empty(),
+            runtime_dir,
+            store,
+            worker_name,
+            parent_scope,
+            persisted_children,
+            valid_records,
+            legacy_projection_exists,
         })
     }
 
@@ -1140,6 +1214,10 @@ impl Drop for SpawnedWorkerRegistry {
             .filter(|record| !record.scope_reclaimed.load(Ordering::Acquire))
             .flat_map(delegated_write_rules)
             .collect::<Vec<_>>();
+        // Empty prepared registries own no delegations, including when abandoned or rejected.
+        if write_rules.is_empty() {
+            return;
+        }
         let _ = parent_scope.update(|current| current.with_removed_deny_rules(write_rules));
     }
 }
@@ -1165,19 +1243,18 @@ fn reclaim_record(
     parent_scope: Option<&SharedScope>,
     record: &SpawnedWorkerRecord,
 ) -> io::Result<()> {
-    if let Ok(path) = worker_allocation::default_allocation_path() {
-        if let Ok(mut guard) = worker_allocation::LockFileGuard::open(&path) {
-            match worker_allocation::reclaim_delegated_scope(
-                &mut guard,
-                parent_name,
-                &record.worker_name,
-                &record.scope_delegated,
-            ) {
-                Ok(()) | Err(worker_allocation::ScopeLockError::UnknownWorker(_)) => {}
-                Err(error) => return Err(io::Error::other(error)),
-            }
-        }
+    let path = worker_allocation::default_allocation_path().map_err(io::Error::other)?;
+    let mut guard = worker_allocation::LockFileGuard::open(&path).map_err(io::Error::other)?;
+    match worker_allocation::reclaim_delegated_scope(
+        &mut guard,
+        parent_name,
+        &record.worker_name,
+        &record.scope_delegated,
+    ) {
+        Ok(()) | Err(worker_allocation::ScopeLockError::UnknownWorker(_)) => {}
+        Err(error) => return Err(io::Error::other(error)),
     }
+    drop(guard);
     if let Some(parent_scope) = parent_scope {
         let write_rules = record
             .scope_delegated
@@ -1240,10 +1317,18 @@ fn store_error_to_io(error: WorkerStoreError) -> io::Error {
 
 #[cfg(test)]
 mod tests {
+    //! Legacy restore cases connect the registry to canonical aggregate storage,
+    //! runtime projections, and temporary SharedScope rules. Successful apply uses
+    //! a corrupt legacy record to exercise durable clearing without consulting the
+    //! process-global legacy allocation file; valid-record scope reclaim is also
+    //! covered by the existing external restore test.
+
+    use std::path::PathBuf;
     use std::time::Duration;
 
     use manifest::{Scope, ScopeConfig};
-    use session_store::LogEntry;
+    use session_store::{LogEntry, WorkerAggregateStore, WorkerMetadata, WorkerSpawnedScopeRule};
+    use tempfile::TempDir;
 
     use super::*;
     use crate::internal_worker::{
@@ -1251,6 +1336,249 @@ mod tests {
         test_internal_worker_session_with_greeting,
     };
 
+    struct LegacyRestoreFixture {
+        runtime_dir: Arc<RuntimeDir>,
+        store: WorkerAggregateStore,
+        parent_scope: SharedScope,
+        scope_root: PathBuf,
+        metadata_path: PathBuf,
+        // Keep all storage and scope paths alive until the other fixture fields are dropped.
+        _root: TempDir,
+    }
+
+    impl LegacyRestoreFixture {
+        async fn new(permission: &str) -> Self {
+            let root = TempDir::new().unwrap();
+            let scope_root = root.path().join("scope");
+            std::fs::create_dir(&scope_root).unwrap();
+            let aggregate = root.path().join("aggregate");
+            let store = WorkerAggregateStore::new(&aggregate, "parent").unwrap();
+            let runtime_dir = Arc::new(
+                RuntimeDir::create(&root.path().join("runtime"), "parent")
+                    .await
+                    .unwrap(),
+            );
+            let mut metadata =
+                WorkerMetadata::new("parent", None).with_workspace_id("restore-test-workspace");
+            metadata.spawned_children.push(WorkerSpawnedChild {
+                worker_name: "legacy-child".into(),
+                socket_path: root.path().join("legacy.sock"),
+                callback_address: runtime_dir.path().join("sock"),
+                scope_delegated: vec![WorkerSpawnedScopeRule {
+                    target: scope_root.clone(),
+                    permission: permission.into(),
+                    recursive: true,
+                    symlink_policy: Default::default(),
+                }],
+            });
+            store.write(&metadata).unwrap();
+            let write_rule = ScopeRule {
+                target: scope_root.clone(),
+                permission: Permission::Write,
+                recursive: true,
+                symlink_policy: Default::default(),
+            };
+            let parent_scope = SharedScope::new(
+                Scope::from_config(&ScopeConfig {
+                    allow: vec![write_rule.clone()],
+                    deny: vec![write_rule],
+                })
+                .unwrap(),
+            );
+            Self {
+                runtime_dir,
+                store,
+                parent_scope,
+                scope_root,
+                metadata_path: aggregate.join("metadata.json"),
+                _root: root,
+            }
+        }
+
+        fn prepare(&self) -> PreparedSpawnedWorkerRegistryLoad<WorkerAggregateStore> {
+            SpawnedWorkerRegistry::prepare_from_worker_state_with_reclaim(
+                self.runtime_dir.clone(),
+                self.store.clone(),
+                "parent".into(),
+                Some(self.parent_scope.clone()),
+            )
+            .unwrap()
+        }
+
+        fn projection_path(&self) -> PathBuf {
+            self.runtime_dir.path().join("spawned_workers.json")
+        }
+    }
+
+    #[tokio::test]
+    async fn registry_preparation_and_abandonment_leave_legacy_state_unchanged() {
+        for projection_exists in [false, true] {
+            let fixture = LegacyRestoreFixture::new("write").await;
+            if projection_exists {
+                // Projection contents are deliberately not decoded during preparation.
+                std::fs::write(fixture.projection_path(), b"obsolete projection").unwrap();
+            }
+            let metadata_before = std::fs::read(&fixture.metadata_path).unwrap();
+            let projection_before = std::fs::read(fixture.projection_path()).ok();
+            let scope_before = fixture.parent_scope.snapshot();
+            assert!(!scope_before.is_writable(&fixture.scope_root));
+
+            let prepared = fixture.prepare();
+
+            assert!(prepared.reclaimed_unreachable);
+            assert!(prepared.registry.list_internal().is_empty());
+            assert!(prepared.registry.internal_worker_snapshots().is_empty());
+            assert_eq!(
+                std::fs::read(&fixture.metadata_path).unwrap(),
+                metadata_before
+            );
+            assert_eq!(
+                std::fs::read(fixture.projection_path()).ok(),
+                projection_before
+            );
+            assert!(Arc::ptr_eq(&scope_before, &fixture.parent_scope.snapshot()));
+            assert!(!fixture.runtime_dir.path().join("sock").exists());
+            drop(prepared);
+            assert_eq!(
+                std::fs::read(&fixture.metadata_path).unwrap(),
+                metadata_before
+            );
+            assert_eq!(
+                std::fs::read(fixture.projection_path()).ok(),
+                projection_before
+            );
+            assert!(Arc::ptr_eq(&scope_before, &fixture.parent_scope.snapshot()));
+        }
+    }
+
+    #[tokio::test]
+    async fn prepared_apply_clears_canonical_children_and_projection_on_same_registry() {
+        let fixture = LegacyRestoreFixture::new("unsupported-legacy-permission").await;
+        std::fs::write(fixture.projection_path(), b"obsolete projection").unwrap();
+        let scope_before = fixture.parent_scope.snapshot();
+        let mut prepared = fixture.prepare();
+        let shared_registry = prepared.registry.clone();
+        // Unrelated metadata may advance while tools are initialized.
+        fixture.store.add_peer("parent", "new-peer").unwrap();
+        let mut expected = fixture.store.read_by_name("parent").unwrap().unwrap();
+        expected.reclaimed_children.extend(
+            expected
+                .spawned_children
+                .iter()
+                .map(reclaimed_child_from_metadata),
+        );
+        expected.spawned_children.clear();
+
+        let loaded = prepared.apply().await.unwrap();
+
+        assert!(loaded.reclaimed_unreachable);
+        assert!(Arc::ptr_eq(&shared_registry, &loaded.registry));
+        assert!(shared_registry.list_internal().is_empty());
+        assert_eq!(
+            fixture.store.read_by_name("parent").unwrap(),
+            Some(expected)
+        );
+        let projection: serde_json::Value =
+            serde_json::from_slice(&std::fs::read(fixture.projection_path()).unwrap()).unwrap();
+        assert_eq!(projection, serde_json::json!([]));
+        // Corrupt records are audited and cleared, not interpreted as scope grants.
+        assert!(Arc::ptr_eq(&scope_before, &fixture.parent_scope.snapshot()));
+        assert!(
+            !fixture
+                .parent_scope
+                .snapshot()
+                .is_writable(&fixture.scope_root)
+        );
+        assert!(!fixture.runtime_dir.path().join("sock").exists());
+    }
+
+    #[tokio::test]
+    async fn prepared_apply_rejects_child_source_drift_before_any_effects() {
+        for drift in ["scope", "socket", "removal", "metadata-deletion"] {
+            let fixture = LegacyRestoreFixture::new("write").await;
+            std::fs::write(fixture.projection_path(), b"obsolete projection").unwrap();
+            let scope_before = fixture.parent_scope.snapshot();
+            let mut prepared = fixture.prepare();
+            if drift == "metadata-deletion" {
+                fixture.store.delete_by_name("parent").unwrap();
+            } else {
+                fixture
+                    .store
+                    .update_by_name("parent", |metadata| match drift {
+                        "scope" => {
+                            metadata.spawned_children[0].scope_delegated[0].recursive = false
+                        }
+                        "socket" => {
+                            metadata.spawned_children[0].socket_path =
+                                fixture.runtime_dir.path().join("replacement.sock");
+                        }
+                        "removal" => metadata.spawned_children.clear(),
+                        _ => unreachable!(),
+                    })
+                    .unwrap();
+            }
+            let metadata_before = std::fs::read(&fixture.metadata_path).ok();
+            let projection_before = std::fs::read(fixture.projection_path()).unwrap();
+
+            let error = prepared
+                .apply()
+                .await
+                .err()
+                .expect("source drift must fail");
+
+            assert_eq!(error.kind(), io::ErrorKind::InvalidData);
+            assert_eq!(std::fs::read(&fixture.metadata_path).ok(), metadata_before);
+            assert_eq!(
+                std::fs::read(fixture.projection_path()).unwrap(),
+                projection_before
+            );
+            assert!(Arc::ptr_eq(&scope_before, &fixture.parent_scope.snapshot()));
+            assert!(!fixture.runtime_dir.path().join("sock").exists());
+        }
+    }
+
+    #[tokio::test]
+    async fn prepared_apply_clears_projection_only_without_reporting_child_reclaim() {
+        let fixture = LegacyRestoreFixture::new("write").await;
+        fixture
+            .store
+            .set_spawned_children("parent", Vec::new())
+            .unwrap();
+        std::fs::write(fixture.projection_path(), b"obsolete projection").unwrap();
+        let metadata_before = std::fs::read(&fixture.metadata_path).unwrap();
+        let mut prepared = fixture.prepare();
+        assert!(!prepared.reclaimed_unreachable);
+
+        let loaded = prepared.apply().await.unwrap();
+
+        assert!(!loaded.reclaimed_unreachable);
+        assert_eq!(
+            std::fs::read(&fixture.metadata_path).unwrap(),
+            metadata_before
+        );
+        let projection: serde_json::Value =
+            serde_json::from_slice(&std::fs::read(fixture.projection_path()).unwrap()).unwrap();
+        assert_eq!(projection, serde_json::json!([]));
+    }
+
+    #[tokio::test]
+    async fn prepared_apply_projection_failure_keeps_canonical_children_and_scope() {
+        let fixture = LegacyRestoreFixture::new("write").await;
+        std::fs::create_dir(fixture.projection_path()).unwrap();
+        let metadata_before = std::fs::read(&fixture.metadata_path).unwrap();
+        let scope_before = fixture.parent_scope.snapshot();
+        let mut prepared = fixture.prepare();
+
+        assert!(prepared.apply().await.is_err());
+
+        assert_eq!(
+            std::fs::read(&fixture.metadata_path).unwrap(),
+            metadata_before
+        );
+        assert!(fixture.projection_path().is_dir());
+        assert!(Arc::ptr_eq(&scope_before, &fixture.parent_scope.snapshot()));
+    }
+
     fn registry() -> Arc<SpawnedWorkerRegistry> {
         let scope = Scope::from_config(&ScopeConfig {
             allow: vec![ScopeRule {

crates/worker/tests/controller_test.rs

@@ -1,3 +1,4 @@
+use std::path::{Path, PathBuf};
 use std::pin::Pin;
 use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
@@ -150,6 +151,19 @@ impl LlmClient for MockClient {
 // Helpers
 // ---------------------------------------------------------------------------
 
+#[path = "../../../tests/support/restore_environment.rs"]
+mod restore_test_environment;
+use restore_test_environment::require_restore_test_environment;
+
+fn owned_relative_runtime_base() -> (tempfile::TempDir, PathBuf) {
+    require_restore_test_environment();
+    let temp = tempfile::tempdir_in(".").unwrap();
+    // The relative reference and the TempDir name the same owned tree, without
+    // putting the checkout prefix into sockaddr_un or changing process cwd.
+    let runtime_base = Path::new(temp.path().file_name().unwrap()).join("runtime");
+    (temp, runtime_base)
+}
+
 fn simple_text_events() -> Vec<LlmEvent> {
     vec![
         LlmEvent::text_block_start(0),
@@ -295,6 +309,7 @@ async fn make_worker_with_manifest_and_workspace_context(
     manifest: WorkerManifest,
     workspace_context: WorkerWorkspaceContext,
 ) -> (Worker<MockClient, TestStore>, std::path::PathBuf) {
+    require_restore_test_environment();
     let store_tmp = tempfile::tempdir().unwrap();
     let store = CombinedStore::new(
         FsStore::new(store_tmp.path()).unwrap(),
@@ -320,18 +335,21 @@ async fn make_worker_with_manifest_and_workspace_context(
 }
 
 async fn spawn_controller(worker: Worker<MockClient, TestStore>) -> WorkerHandle {
-    let tmp = tempfile::tempdir().unwrap();
-    let runtime_base = tmp.path().to_owned();
-    std::mem::forget(tmp);
-    let (handle, _shutdown_rx) = WorkerController::spawn(worker, &runtime_base)
+    let (tmp, runtime_base) = owned_relative_runtime_base();
+    let (handle, shutdown_rx) = WorkerController::spawn(worker, &runtime_base)
         .await
         .unwrap();
+    // Retain ownership until controller shutdown (or test-runtime teardown).
+    tokio::spawn(async move {
+        let _ = shutdown_rx.await;
+        drop(tmp);
+    });
     handle
 }
 
 #[tokio::test]
 async fn canonical_worker_handle_uploads_and_submits_attachment() {
-    let root = tempfile::tempdir().unwrap();
+    let (root, runtime_base) = owned_relative_runtime_base();
     let manifest = WorkerManifest::from_toml(MANIFEST_TOML).unwrap();
     let session = session_store::WorkerSessionStore::new(root.path().join("session")).unwrap();
     let store = CombinedStore::new(
@@ -353,7 +371,7 @@ async fn canonical_worker_handle_uploads_and_submits_attachment() {
     )
     .await
     .unwrap();
-    let (handle, shutdown_rx) = WorkerController::spawn(worker, &root.path().join("runtime"))
+    let (handle, shutdown_rx) = WorkerController::spawn(worker, &runtime_base)
         .await
         .unwrap();
     let mut rx = handle.subscribe();
@@ -762,6 +780,227 @@ async fn controller_startup_failure_closes_bound_workdir_session() {
     ));
 }
 
+mod registry_startup_failure {
+    //! Connect real controller preparation, registry source fences and session cleanup.
+    //! The Workdir telemetry boundary substitutes a concurrent metadata change at subscription;
+    //! the real registry must reject drift before starting its prepared observer.
+    use super::*;
+    use session_store::WorkerMetadataStore;
+    use tokio::sync::broadcast;
+    use workdir::WorkdirSession;
+
+    struct ObservedSession {
+        local: LocalWorkdirSession,
+        metadata_store: TestStore,
+        worker_name: String,
+        events: broadcast::Sender<workdir::CommandEvent>,
+        subscriptions: AtomicUsize,
+        receivers_at_close: Mutex<Vec<usize>>,
+    }
+
+    impl std::fmt::Debug for ObservedSession {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            f.debug_struct("ObservedSession")
+                .field("worker_name", &self.worker_name)
+                .finish_non_exhaustive()
+        }
+    }
+
+    #[async_trait]
+    impl WorkdirSession for ObservedSession {
+        fn workdir(&self) -> &Workdir {
+            self.local.workdir()
+        }
+
+        fn capabilities(&self) -> WorkdirSessionCapabilities {
+            self.local.capabilities()
+        }
+
+        async fn stat(
+            &self,
+            request: workdir::StatRequest,
+        ) -> Result<workdir::StatResult, WorkdirError> {
+            self.local.stat(request).await
+        }
+
+        async fn read(
+            &self,
+            request: workdir::ReadRequest,
+        ) -> Result<workdir::ReadResult, WorkdirError> {
+            self.local.read(request).await
+        }
+
+        async fn write(
+            &self,
+            request: workdir::WriteRequest,
+        ) -> Result<workdir::WriteResult, WorkdirError> {
+            self.local.write(request).await
+        }
+
+        async fn edit(
+            &self,
+            request: workdir::EditRequest,
+        ) -> Result<workdir::EditResult, WorkdirError> {
+            self.local.edit(request).await
+        }
+
+        async fn list(
+            &self,
+            request: workdir::ListRequest,
+        ) -> Result<workdir::ListResult, WorkdirError> {
+            self.local.list(request).await
+        }
+
+        async fn glob(
+            &self,
+            request: workdir::GlobRequest,
+        ) -> Result<workdir::GlobResult, WorkdirError> {
+            self.local.glob(request).await
+        }
+
+        async fn grep(
+            &self,
+            request: workdir::GrepRequest,
+        ) -> Result<workdir::GrepResult, WorkdirError> {
+            self.local.grep(request).await
+        }
+
+        async fn start_command(
+            &self,
+            request: CommandRequest,
+        ) -> Result<workdir::CommandHandle, WorkdirError> {
+            self.local.start_command(request).await
+        }
+
+        async fn command_status(
+            &self,
+            handle: workdir::CommandHandle,
+        ) -> Result<workdir::CommandStatus, WorkdirError> {
+            self.local.command_status(handle).await
+        }
+
+        async fn command_output(
+            &self,
+            request: CommandOutputRequest,
+        ) -> Result<workdir::CommandOutput, WorkdirError> {
+            self.local.command_output(request).await
+        }
+
+        async fn cancel_command(&self, handle: workdir::CommandHandle) -> Result<(), WorkdirError> {
+            self.local.cancel_command(handle).await
+        }
+
+        fn subscribe_command_events(&self) -> Option<broadcast::Receiver<workdir::CommandEvent>> {
+            // Interleave a canonical change after read-only registry preparation.
+            let mut metadata = self
+                .metadata_store
+                .read_by_name(&self.worker_name)
+                .unwrap()
+                .unwrap_or_else(|| session_store::WorkerMetadata::new(&self.worker_name, None));
+            metadata
+                .spawned_children
+                .push(session_store::WorkerSpawnedChild {
+                    worker_name: "legacy-drift".into(),
+                    socket_path: PathBuf::new(),
+                    callback_address: PathBuf::new(),
+                    scope_delegated: vec![],
+                });
+            self.metadata_store.write(&metadata).unwrap();
+            self.subscriptions.fetch_add(1, Ordering::SeqCst);
+            Some(self.events.subscribe())
+        }
+
+        async fn close(&self) -> Result<(), WorkdirError> {
+            // Record before any await: aborting without joining must not pass by getting a
+            // scheduler turn during the provider's asynchronous close. The sender stays alive
+            // even after close, so channel closure cannot accidentally dispose of the observer.
+            self.receivers_at_close
+                .lock()
+                .unwrap()
+                .push(self.events.receiver_count());
+            self.local.close().await
+        }
+    }
+
+    #[tokio::test(flavor = "current_thread")]
+    async fn registry_drift_drops_prepared_observer_before_closing_session() {
+        let root = tempfile::tempdir().unwrap();
+        let run_dir = root.path().join("runs/1");
+        let client = MockClient::new(simple_text_events());
+        let manifest = WorkerManifest::from_toml(MANIFEST_TOML).unwrap();
+        let store = CombinedStore::new(
+            FsStore::new(root.path().join("sessions")).unwrap(),
+            FsWorkerStore::new(root.path().join("workers")).unwrap(),
+        );
+        let engine =
+            Engine::<_, agen::state::Mutable, worker::SessionHistoryMetadata>::new_annotated(
+                client.clone(),
+            );
+        let mut worker = Worker::new(
+            manifest,
+            engine,
+            store.clone(),
+            WorkerWorkspaceContext::local_filesystem(None),
+            WorkerFilesystemAuthority::local(root.path().into(), root.path().into()),
+            manifest::Scope::writable(root.path()).unwrap(),
+        )
+        .await
+        .unwrap();
+        let (events, unused_receiver) = broadcast::channel(8);
+        drop(unused_receiver);
+        let session = Arc::new(ObservedSession {
+            local: LocalWorkdirSession::materialized_bound(
+                Workdir::new("registry-startup-failure"),
+                root.path().into(),
+                root.path().into(),
+                worker.scope().clone(),
+                WorkdirSessionCapabilities::ALL,
+            ),
+            metadata_store: store,
+            worker_name: worker.manifest().worker.name.clone(),
+            events,
+            subscriptions: AtomicUsize::new(0),
+            receivers_at_close: Mutex::new(Vec::new()),
+        });
+        worker.bind_single_workdir_session(Some(session.clone()));
+        let sessions = worker.workdir_sessions();
+
+        let result = tokio::time::timeout(
+            std::time::Duration::from_secs(5),
+            WorkerController::spawn_runtime_managed_run(worker, &run_dir),
+        )
+        .await
+        .expect("registry rejection and cleanup must settle");
+        let error = match result {
+            Err(error) => error,
+            Ok(_) => panic!("registry source drift must reject controller startup"),
+        };
+        assert_eq!(error.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(session.subscriptions.load(Ordering::SeqCst), 1);
+        assert_eq!(*session.receivers_at_close.lock().unwrap(), vec![0]);
+        assert_eq!(session.events.receiver_count(), 0);
+        assert!(sessions.aliases().is_empty());
+        // No controller retains the Worker/Engine after the failed startup.
+        assert_eq!(Arc::strong_count(&client.call_count), 1);
+        assert_eq!(client.call_count.load(Ordering::SeqCst), 0);
+        assert!(!run_dir.join("worker.sock").exists());
+        assert!(!run_dir.join("status.json").exists());
+        assert!(!run_dir.join("manifest.toml").exists());
+        assert!(matches!(
+            session
+                .start_command(CommandRequest {
+                    command: "printf unreachable".to_owned(),
+                    timeout_secs: 5,
+                    transport_byte_budget: workdir::MAX_COMMAND_TRANSPORT_BYTES,
+                    cwd: workdir::WorkdirPath::root(),
+                    tool_call_id: None,
+                })
+                .await,
+            Err(ref error) if matches!(error.classification_source(), WorkdirError::Unavailable(_))
+        ));
+    }
+}
+
 async fn wait_for_status(handle: &WorkerHandle, status: WorkerStatus) {
     let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(2);
     loop {
@@ -1561,11 +1800,12 @@ permission = "write"
 "#;
     let client = MockClient::new(simple_text_events());
     let worker = make_worker_with_pwd_and_manifest(client, manifest).await.0;
-    let tmp = tempfile::tempdir().unwrap();
-    let result = WorkerController::spawn(worker, tmp.path()).await;
+    let (_tmp, runtime_base) = owned_relative_runtime_base();
+    let result = WorkerController::spawn(worker, &runtime_base).await;
     assert!(
         result.is_ok(),
-        "feature exposure must not imply delegation authority"
+        "feature exposure must not imply delegation authority: {:?}",
+        result.as_ref().err()
     );
 }
 

crates/workspace-server/Cargo.toml

@@ -10,6 +10,10 @@ autobins = false
 name = "yoi-server"
 path = "src/main.rs"
 
+[features]
+# Offline fixture preparation only; build product binaries without this feature.
+test-support = []
+
 [dependencies]
 api-macros = { workspace = true, features = ["axum"] }
 async-trait.workspace = true

crates/workspace-server/src/hosts.rs

@@ -5113,6 +5113,9 @@ impl WorkspaceWorkerRuntime for RemoteWorkerRuntime {
                         runtime_api::WorkerRestoreState::RolledBack => {
                             server_api::WorkerRestoreState::RolledBack
                         }
+                        runtime_api::WorkerRestoreState::Superseded => {
+                            server_api::WorkerRestoreState::Superseded
+                        }
                         runtime_api::WorkerRestoreState::ReconciliationRequired => {
                             server_api::WorkerRestoreState::ReconciliationRequired
                         }

crates/workspace-server/src/lib.rs

@@ -19,12 +19,17 @@ pub mod repositories;
 pub mod repository_access;
 pub mod repository_source;
 pub mod resource_broker;
+#[cfg(test)]
+#[path = "../../../tests/support/restore_environment.rs"]
+mod restore_test_environment;
 pub mod retention;
 pub mod runtime_settings;
 pub mod runtime_subscription;
 pub mod server;
 pub mod skills;
 pub mod store;
+#[cfg(feature = "test-support")]
+pub use store::fixture_support as test_support;
 pub mod subjektiv;
 #[cfg(test)]
 mod ticket_browse_performance_tests;

crates/workspace-server/src/server.rs

@@ -48742,6 +48742,12 @@ mod tests {
             std::sync::Mutex<Vec<worker_runtime::execution::WorkerLifecycleOperationId>>,
         reconcile_operations:
             std::sync::Mutex<Vec<worker_runtime::execution::WorkerLifecycleOperationId>>,
+        restore_owners: std::sync::Mutex<
+            std::collections::HashMap<
+                worker_runtime::execution::WorkerLifecycleOperationId,
+                worker_runtime::identity::WorkerRef,
+            >,
+        >,
         inputs: std::sync::Mutex<Vec<(worker_runtime::identity::WorkerRef, String)>>,
         protocol_methods:
             Arc<std::sync::Mutex<Vec<(worker_runtime::identity::WorkerRef, protocol::Method)>>>,
@@ -48775,6 +48781,7 @@ mod tests {
                 restore_failure: std::sync::Mutex::new(None),
                 restore_operations: std::sync::Mutex::new(Vec::new()),
                 reconcile_operations: std::sync::Mutex::new(Vec::new()),
+                restore_owners: Default::default(),
                 inputs: std::sync::Mutex::new(Vec::new()),
                 protocol_methods: Arc::new(std::sync::Mutex::new(Vec::new())),
                 stops: std::sync::Mutex::new(Vec::new()),
@@ -48979,6 +48986,10 @@ mod tests {
                 .lock()
                 .unwrap()
                 .push(request.operation_id);
+            self.restore_owners
+                .lock()
+                .unwrap()
+                .insert(request.operation_id, request.worker_ref.clone());
             if let Some(result) = self.restore_failure.lock().unwrap().take() {
                 return result;
             }
@@ -49014,6 +49025,21 @@ mod tests {
             }
         }
 
+        fn restore_recovery_disposition(
+            &self,
+            operation_id: worker_runtime::execution::WorkerLifecycleOperationId,
+            worker_ref: &worker_runtime::identity::WorkerRef,
+        ) -> Option<worker_runtime::execution::RestoreRecoveryDisposition> {
+            use worker_runtime::execution::RestoreRecoveryDisposition as D;
+            Some(
+                if self.restore_owners.lock().unwrap().get(&operation_id) == Some(worker_ref) {
+                    D::Continue
+                } else {
+                    D::Absent
+                },
+            )
+        }
+
         fn reconcile_restore(
             &self,
             request: worker_runtime::execution::WorkerExecutionRestoreRequest,

crates/workspace-server/src/server/tests/restored_parent_control_tests.rs

@@ -11,6 +11,20 @@ use worker_runtime::worker_backend::{
     WorkerRuntimeExecutionBackend,
 };
 
+pub(super) fn connected_restore_tempdir() -> tempfile::TempDir {
+    crate::restore_test_environment::require_restore_test_environment();
+    let checkout = Path::new(env!("CARGO_MANIFEST_DIR"))
+        .parent()
+        .unwrap()
+        .parent()
+        .unwrap();
+    tempfile::Builder::new()
+        .prefix("c")
+        .rand_bytes(3)
+        .tempdir_in(checkout.join("target"))
+        .unwrap()
+}
+
 const WATCHDOG: Duration = Duration::from_secs(45);
 
 #[derive(Clone)]
@@ -82,24 +96,91 @@ impl RuntimeWorkerFactory for CapturingFactory {
     ) -> std::result::Result<(), String> {
         self.inner.preflight_restore(request).await
     }
-    async fn restore_controller(
+    async fn recover_restore_source(
+        &self,
+        request: &worker_runtime::execution::WorkerExecutionRestoreRequest,
+    ) -> std::result::Result<bool, worker_runtime::worker_backend::RuntimeRestoreFailure> {
+        self.inner.recover_restore_source(request).await
+    }
+
+    async fn prepare_restore(
         &self,
         request: worker_runtime::execution::WorkerExecutionRestoreRequest,
-    ) -> std::result::Result<RuntimeWorkerController, String> {
+    ) -> std::result::Result<
+        Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
         let worker_ref = request.worker_ref.clone();
-        let controller = self.inner.restore_controller(request).await?;
+        let inner = self.inner.prepare_restore(request).await?;
+        Ok(Box::new(CapturingCandidate {
+            inner,
+            worker_ref,
+            captures: self.captures.clone(),
+        }))
+    }
+
+    async fn cleanup_failed_restore(
+        &self,
+        request: &worker_runtime::execution::WorkerExecutionRestoreRequest,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inner.cleanup_failed_restore(request).await
+    }
+
+    async fn reconcile_stopped_worker(
+        &self,
+        worker_ref: &worker_runtime::identity::WorkerRef,
+    ) -> std::result::Result<(), String> {
+        self.inner.reconcile_stopped_worker(worker_ref).await
+    }
+
+    fn activate_restored_controller(
+        &self,
+        worker_ref: &worker_runtime::identity::WorkerRef,
+        workspace_id: Option<&str>,
+        handle: &worker::WorkerHandle,
+    ) -> std::result::Result<(), String> {
+        self.inner
+            .activate_restored_controller(worker_ref, workspace_id, handle)
+    }
+}
+
+struct CapturingCandidate {
+    inner: Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+    worker_ref: worker_runtime::identity::WorkerRef,
+    captures: Arc<Captures>,
+}
+
+#[async_trait::async_trait]
+impl worker_runtime::worker_backend::RuntimeRestoreCandidate for CapturingCandidate {
+    fn initial_state(&self) -> protocol::WorkerStateSnapshot {
+        self.inner.initial_state()
+    }
+
+    async fn activate(
+        &mut self,
+    ) -> std::result::Result<
+        RuntimeWorkerController,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        let controller = self.inner.activate().await?;
         self.captures
             .handles
             .lock()
             .unwrap()
-            .push((worker_ref.clone(), controller.handle.clone()));
+            .push((self.worker_ref.clone(), controller.handle.clone()));
         self.captures
             .restored_clients
             .lock()
             .unwrap()
-            .push((worker_ref, controller.workspace_client.clone()));
+            .push((self.worker_ref.clone(), controller.workspace_client.clone()));
         Ok(controller)
     }
+
+    async fn discard(
+        &mut self,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inner.discard().await
+    }
 }
 
 struct Resources {
@@ -259,7 +340,7 @@ async fn call(tools: &BTreeMap<String, Arc<dyn Tool>>, name: &str, args: Value)
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
 async fn restored_runtime_parent_lists_and_inputs_existing_child() {
-    let root = tempfile::tempdir().unwrap();
+    let root = connected_restore_tempdir();
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
     let base = format!("http://{}", listener.local_addr().unwrap());
     let mut config = test_server_config(root.path());

crates/workspace-server/src/server_assignment_attention_tests.rs

@@ -476,6 +476,7 @@ impl AttentionWorkerFactory {
             shutdown: Arc::new(tokio::sync::Mutex::new(Some(started.shutdown))),
             controller_task: started.controller_task,
             workspace_client,
+            restore_streams: None,
         })
     }
 }
@@ -489,12 +490,61 @@ impl worker_runtime::worker_backend::RuntimeWorkerFactory for AttentionWorkerFac
         self.controller(request.worker_ref.worker_id.to_string(), false)
             .await
     }
-    async fn restore_controller(
+    async fn prepare_restore(
         &self,
         request: worker_runtime::execution::WorkerExecutionRestoreRequest,
-    ) -> std::result::Result<worker_runtime::worker_backend::RuntimeWorkerController, String> {
-        self.controller(request.worker_ref.worker_id.to_string(), true)
-            .await
+    ) -> std::result::Result<
+        Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        Ok(Box::new(AttentionRestoreCandidate {
+            inputs: Some((
+                AttentionWorkerFactory {
+                    root: self.root.clone(),
+                },
+                request.worker_ref.worker_id.to_string(),
+            )),
+        }))
+    }
+}
+
+// Preparation owns only fixture inputs; reconstruction and Controller startup
+// remain lazy so discard never needs to shut down a started Worker.
+struct AttentionRestoreCandidate {
+    inputs: Option<(AttentionWorkerFactory, String)>,
+}
+
+#[async_trait::async_trait]
+impl worker_runtime::worker_backend::RuntimeRestoreCandidate for AttentionRestoreCandidate {
+    fn initial_state(&self) -> protocol::WorkerStateSnapshot {
+        protocol::WorkerStateSnapshot::initial()
+    }
+
+    async fn activate(
+        &mut self,
+    ) -> std::result::Result<
+        worker_runtime::worker_backend::RuntimeWorkerController,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        let (factory, worker_id) = self.inputs.take().ok_or_else(|| {
+            worker_runtime::worker_backend::RuntimeRestoreFailure::refused(
+                worker_runtime::worker_backend::RuntimeRestoreStage::Controller,
+                "attention fixture candidate already consumed",
+            )
+        })?;
+        factory.controller(worker_id, true).await.map_err(|_| {
+            worker_runtime::worker_backend::RuntimeRestoreFailure::unproven(
+                worker_runtime::worker_backend::RuntimeRestoreStage::Reconstruction,
+                "attention fixture reconstruction failed",
+            )
+        })
+    }
+
+    async fn discard(
+        &mut self,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inputs.take();
+        Ok(())
     }
 }
 
@@ -601,7 +651,7 @@ async fn orchestrator_attention_real_worker_lost_response_retry_commits_once() {
 
 async fn real_attention_worker_start_restore(lose_response: bool) {
     use session_store::{LogEntry, Store, WorkerSessionStore};
-    let dir = tempfile::tempdir().unwrap();
+    let dir = restored_parent_control_tests::connected_restore_tempdir();
     init_clean_git_workspace(dir.path());
     let config = test_server_config(dir.path());
     let store = Arc::new(SqliteWorkspaceStore::open_for_test(&config.database_path).unwrap());

crates/workspace-server/src/server_assignment_delivery_tests.rs

@@ -14,17 +14,37 @@ async fn notification_diagnostics_http_auth_scope_and_readonly_evidence() {
     seed_test_api_token(api.store.as_ref(), TEST_WORKSPACE_ID);
     let ticket = delivery_ticket(&api, "Notification diagnostics auth fixture");
     delivery_assign(&api, ticket.id.as_str(), "diagnostics-assignment").await;
-    let path = format!("/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?ticket={}&limit=1",ticket.id.as_str());
-    let page = delivery_http(&api,"GET",&path,None,StatusCode::OK).await;
-    assert_eq!(page["items"].as_array().unwrap().len(),1);
-    assert_eq!(page["items"][0]["model_processing_confirmed"],false);
+    let path = format!(
+        "/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?ticket={}&limit=1",
+        ticket.id.as_str()
+    );
+    let page = delivery_http(&api, "GET", &path, None, StatusCode::OK).await;
+    assert_eq!(page["items"].as_array().unwrap().len(), 1);
+    assert_eq!(page["items"][0]["model_processing_confirmed"], false);
     assert!(page["items"][0].get("content").is_none());
-    let again=delivery_http(&api,"GET",&path,None,StatusCode::OK).await;
-    assert_eq!(again,page);
-    let response=build_router(api.clone()).oneshot(Request::builder().uri(&path).body(Body::empty()).unwrap()).await.unwrap();
-    assert_eq!(response.status(),StatusCode::UNAUTHORIZED);
-    delivery_http(&api,"GET","/api/w/other-workspace/notification-diagnostics",None,StatusCode::NOT_FOUND).await;
-    delivery_http(&api,"GET",&format!("/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?limit=101"),None,StatusCode::BAD_REQUEST).await;
+    let again = delivery_http(&api, "GET", &path, None, StatusCode::OK).await;
+    assert_eq!(again, page);
+    let response = build_router(api.clone())
+        .oneshot(Request::builder().uri(&path).body(Body::empty()).unwrap())
+        .await
+        .unwrap();
+    assert_eq!(response.status(), StatusCode::UNAUTHORIZED);
+    delivery_http(
+        &api,
+        "GET",
+        "/api/w/other-workspace/notification-diagnostics",
+        None,
+        StatusCode::NOT_FOUND,
+    )
+    .await;
+    delivery_http(
+        &api,
+        "GET",
+        &format!("/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?limit=101"),
+        None,
+        StatusCode::BAD_REQUEST,
+    )
+    .await;
 }
 
 fn delivery_assignment_path(ticket: &str) -> String {
@@ -814,6 +834,7 @@ impl DeliveryWorkerFactory {
             shutdown: Arc::new(tokio::sync::Mutex::new(Some(started.shutdown))),
             controller_task: started.controller_task,
             workspace_client,
+            restore_streams: None,
         })
     }
 }
@@ -827,12 +848,62 @@ impl worker_runtime::worker_backend::RuntimeWorkerFactory for DeliveryWorkerFact
         self.controller(request.worker_ref.worker_id.to_string(), false)
             .await
     }
-    async fn restore_controller(
+    async fn prepare_restore(
         &self,
         request: delivery_execution::WorkerExecutionRestoreRequest,
-    ) -> std::result::Result<worker_runtime::worker_backend::RuntimeWorkerController, String> {
-        self.controller(request.worker_ref.worker_id.to_string(), true)
-            .await
+    ) -> std::result::Result<
+        Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        Ok(Box::new(DeliveryRestoreCandidate {
+            inputs: Some((
+                DeliveryWorkerFactory {
+                    root: self.root.clone(),
+                    client: self.client.clone(),
+                },
+                request.worker_ref.worker_id.to_string(),
+            )),
+        }))
+    }
+}
+
+// Retain the same scripted client/receipt store without opening either store or
+// starting a Controller until activation. Discard drops only these owned inputs.
+struct DeliveryRestoreCandidate {
+    inputs: Option<(DeliveryWorkerFactory, String)>,
+}
+
+#[async_trait::async_trait]
+impl worker_runtime::worker_backend::RuntimeRestoreCandidate for DeliveryRestoreCandidate {
+    fn initial_state(&self) -> protocol::WorkerStateSnapshot {
+        protocol::WorkerStateSnapshot::initial()
+    }
+
+    async fn activate(
+        &mut self,
+    ) -> std::result::Result<
+        worker_runtime::worker_backend::RuntimeWorkerController,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        let (factory, worker_id) = self.inputs.take().ok_or_else(|| {
+            worker_runtime::worker_backend::RuntimeRestoreFailure::refused(
+                worker_runtime::worker_backend::RuntimeRestoreStage::Controller,
+                "delivery fixture candidate already consumed",
+            )
+        })?;
+        factory.controller(worker_id, true).await.map_err(|_| {
+            worker_runtime::worker_backend::RuntimeRestoreFailure::unproven(
+                worker_runtime::worker_backend::RuntimeRestoreStage::Reconstruction,
+                "delivery fixture reconstruction failed",
+            )
+        })
+    }
+
+    async fn discard(
+        &mut self,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inputs.take();
+        Ok(())
     }
 }
 
@@ -931,6 +1002,7 @@ async fn delivery_real_fixture(
     Arc<DeliveryResponseLossBackend>,
     DeliveryScriptedClient,
 ) {
+    let _environment = restored_parent_control_tests::connected_restore_tempdir();
     let config = test_server_config(root);
     let store = Arc::new(SqliteWorkspaceStore::open_for_test(&config.database_path).unwrap());
     seed_test_registered_workspace(store.as_ref(), &config)
@@ -1136,9 +1208,22 @@ async fn assignment_real_worker_response_loss_restore_and_concurrent_scans_commi
         "uncertain delivery changed request id or payload: {attempts:?}"
     );
     assert_eq!(delivery_receipts(&api, &worker).len(), 1);
-    let unknown = api.store.notification_diagnostics(TEST_WORKSPACE_ID, &server_api::NotificationDiagnosticsQuery {
-        assignment_id: Some(assignment.into()), ..Default::default()
-    }).unwrap().items.into_iter().find(|item| item.stage == server_api::NotificationDiagnosticStage::AcceptanceUnknown).expect("saved unknown acceptance must remain visible while Runtime already has its receipt");
+    let unknown = api
+        .store
+        .notification_diagnostics(
+            TEST_WORKSPACE_ID,
+            &server_api::NotificationDiagnosticsQuery {
+                assignment_id: Some(assignment.into()),
+                ..Default::default()
+            },
+        )
+        .unwrap()
+        .items
+        .into_iter()
+        .find(|item| item.stage == server_api::NotificationDiagnosticStage::AcceptanceUnknown)
+        .expect(
+            "saved unknown acceptance must remain visible while Runtime already has its receipt",
+        );
     assert!(!unknown.model_processing_confirmed);
     // Change mutable rendering context while the acknowledgement is unknown.
     // A new self-report is not a new input and cannot change the frozen retry.
@@ -1199,11 +1284,24 @@ async fn assignment_real_worker_response_loss_restore_and_concurrent_scans_commi
     }
     delivery_wait_idle(&reopened, &worker, 2).await;
     assert_eq!(delivery_receipts(&reopened, &worker).len(), 2);
-    let evidence = delivery_http(&reopened,"GET",&format!("/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?source_id={}",unknown.source_id),None,StatusCode::OK).await;
-    assert_eq!(evidence["items"][0]["stage"],"runtime_accepted");
-    assert_eq!(evidence["items"][0]["delivery_id"],json!(unknown.delivery_id));
-    assert_eq!(evidence["items"][0]["model_processing_confirmed"],false);
-    assert_eq!(evidence["items"][0]["retry_candidate"],false);
+    let evidence = delivery_http(
+        &reopened,
+        "GET",
+        &format!(
+            "/api/w/{TEST_WORKSPACE_ID}/notification-diagnostics?source_id={}",
+            unknown.source_id
+        ),
+        None,
+        StatusCode::OK,
+    )
+    .await;
+    assert_eq!(evidence["items"][0]["stage"], "runtime_accepted");
+    assert_eq!(
+        evidence["items"][0]["delivery_id"],
+        json!(unknown.delivery_id)
+    );
+    assert_eq!(evidence["items"][0]["model_processing_confirmed"], false);
+    assert_eq!(evidence["items"][0]["retry_candidate"], false);
     assert_eq!(
         delivery_commits(&reopened, &worker)
             .iter()

crates/workspace-server/src/server_workspace_config_wip_tests.rs

@@ -187,7 +187,14 @@ async fn config_call(
     runtime
         .invoke(
             path.into(),
-            if matches!(operation,"edit"|"write") {wip_protocol::InterfaceReference {scope:"/".into(),name:format!("yoi.file/{operation}-text/v1")}} else {config_interface(path)},
+            if matches!(operation, "edit" | "write") {
+                wip_protocol::InterfaceReference {
+                    scope: "/".into(),
+                    name: format!("yoi.file/{operation}-text/v1"),
+                }
+            } else {
+                config_interface(path)
+            },
             operation.into(),
             args,
             agen::tool::ToolExecutionContext::direct(),
@@ -581,11 +588,88 @@ impl worker_runtime::worker_backend::RuntimeWorkerFactory for ConfigCaptureFacto
         *self.controller_abort.lock().unwrap() = Some(controller.controller_task.abort_handle());
         Ok(controller)
     }
-    async fn restore_controller(
+    async fn preflight_restore(
+        &self,
+        request: &worker_runtime::execution::WorkerExecutionRestoreRequest,
+    ) -> std::result::Result<(), String> {
+        self.inner.preflight_restore(request).await
+    }
+
+    async fn recover_restore_source(
+        &self,
+        request: &worker_runtime::execution::WorkerExecutionRestoreRequest,
+    ) -> std::result::Result<bool, worker_runtime::worker_backend::RuntimeRestoreFailure> {
+        self.inner.recover_restore_source(request).await
+    }
+
+    async fn prepare_restore(
         &self,
         request: worker_runtime::execution::WorkerExecutionRestoreRequest,
-    ) -> std::result::Result<worker_runtime::worker_backend::RuntimeWorkerController, String> {
-        self.inner.restore_controller(request).await
+    ) -> std::result::Result<
+        Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        let inner = self.inner.prepare_restore(request).await?;
+        Ok(Box::new(ConfigCaptureCandidate {
+            inner,
+            client: self.client.clone(),
+            controller_abort: self.controller_abort.clone(),
+        }))
+    }
+
+    async fn cleanup_failed_restore(
+        &self,
+        request: &worker_runtime::execution::WorkerExecutionRestoreRequest,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inner.cleanup_failed_restore(request).await
+    }
+
+    async fn reconcile_stopped_worker(
+        &self,
+        worker_ref: &worker_runtime::identity::WorkerRef,
+    ) -> std::result::Result<(), String> {
+        self.inner.reconcile_stopped_worker(worker_ref).await
+    }
+
+    fn activate_restored_controller(
+        &self,
+        worker_ref: &worker_runtime::identity::WorkerRef,
+        workspace_id: Option<&str>,
+        handle: &worker::WorkerHandle,
+    ) -> std::result::Result<(), String> {
+        self.inner
+            .activate_restored_controller(worker_ref, workspace_id, handle)
+    }
+}
+
+struct ConfigCaptureCandidate {
+    inner: Box<dyn worker_runtime::worker_backend::RuntimeRestoreCandidate>,
+    client: Arc<Mutex<Option<Arc<dyn worker::WorkspaceClient>>>>,
+    controller_abort: Arc<Mutex<Option<tokio::task::AbortHandle>>>,
+}
+
+#[async_trait]
+impl worker_runtime::worker_backend::RuntimeRestoreCandidate for ConfigCaptureCandidate {
+    fn initial_state(&self) -> protocol::WorkerStateSnapshot {
+        self.inner.initial_state()
+    }
+
+    async fn activate(
+        &mut self,
+    ) -> std::result::Result<
+        worker_runtime::worker_backend::RuntimeWorkerController,
+        worker_runtime::worker_backend::RuntimeRestoreFailure,
+    > {
+        let controller = self.inner.activate().await?;
+        *self.client.lock().unwrap() = Some(controller.workspace_client.clone());
+        *self.controller_abort.lock().unwrap() = Some(controller.controller_task.abort_handle());
+        Ok(controller)
+    }
+
+    async fn discard(
+        &mut self,
+    ) -> std::result::Result<(), worker_runtime::worker_backend::RuntimeRestoreCleanupFailure> {
+        self.inner.discard().await
     }
 }
 
@@ -714,6 +798,7 @@ async fn workspace_config_actual_remote_and_embedded_clients_use_authenticated_w
         RUNTIME_REQUEST_SOURCE_PROOF_HEADER, WORKSPACE_REQUEST_PERMISSION,
         decode_runtime_request_source_claims,
     };
+    let root = restored_parent_control_tests::connected_restore_tempdir();
     // Remote Runtime-owned HTTP adapter, including its real request signer.
     let mut fixture = manual_worker_assignment_fixture().await;
     let identity = RuntimeIdentityMaterial::generate(&fixture.worker.runtime_id).unwrap();
@@ -793,7 +878,6 @@ async fn workspace_config_actual_remote_and_embedded_clients_use_authenticated_w
 
     // Actual embedded Runtime factory -> real Worker controller -> captured
     // production WorkspaceClient -> signed HTTP -> production Backend middleware.
-    let root = tempfile::tempdir().unwrap();
     init_clean_git_workspace(root.path());
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
     let addr = listener.local_addr().unwrap();

crates/workspace-server/src/store.rs

@@ -2603,6 +2603,10 @@ pub struct SqliteWorkspaceStore {
     db_executor: crate::db_executor::DbExecutor,
 }
 
+#[cfg(feature = "test-support")]
+#[path = "test_support.rs"]
+pub mod fixture_support;
+
 impl SqliteWorkspaceStore {
     pub fn open(path: impl AsRef<Path>) -> Result<Self> {
         let conn = Connection::open(path)?;
@@ -2914,6 +2918,33 @@ impl SqliteWorkspaceStore {
         request_fingerprint: &str,
         singleton_key: Option<&str>,
         current_memory_settings: &WorkspaceMemorySettingsRecord,
+    ) -> Result<WorkerCreateReservation> {
+        self.reserve_worker_create_with(
+            workspace_id,
+            runtime_id,
+            allocation_key,
+            request_fingerprint,
+            singleton_key,
+            current_memory_settings,
+            |_| Ok(None),
+            |_, _| Ok(()),
+        )
+    }
+
+    // Keep reservation validation, Memory binding and writes in one transaction. The
+    // nondefault test-support module supplies fixed identity / offline catalog hooks;
+    // ordinary creation always supplies a newly generated ID and no completion hook.
+    #[allow(clippy::too_many_arguments)]
+    fn reserve_worker_create_with(
+        &self,
+        workspace_id: &str,
+        runtime_id: &str,
+        allocation_key: &str,
+        request_fingerprint: &str,
+        singleton_key: Option<&str>,
+        current_memory_settings: &WorkspaceMemorySettingsRecord,
+        prepare: impl FnOnce(&rusqlite::Transaction<'_>) -> Result<Option<WorkerId>>,
+        complete: impl FnOnce(&rusqlite::Transaction<'_>, &WorkerCreateReservation) -> Result<()>,
     ) -> Result<WorkerCreateReservation> {
         if allocation_key.trim().is_empty() || request_fingerprint.trim().is_empty() {
             return Err(Error::InvalidInput(
@@ -2926,6 +2957,7 @@ impl SqliteWorkspaceStore {
         validate_workspace_memory_settings_record(current_memory_settings, workspace_id)?;
         self.with_conn_mut(|conn| {
             let tx = conn.transaction_with_behavior(TransactionBehavior::Immediate)?;
+            let fixed_worker_id = prepare(&tx)?;
             let existing = tx
                 .query_row(
                     "SELECT worker_id, runtime_id, request_fingerprint, create_fingerprint, \
@@ -3108,7 +3140,7 @@ impl SqliteWorkspaceStore {
             validate_workspace_memory_settings_snapshot(&snapshot, workspace_id)?;
             let create_fingerprint =
                 bound_worker_create_fingerprint(request_fingerprint, &snapshot);
-            let worker_id = WorkerId::now_v7();
+            let worker_id = fixed_worker_id.unwrap_or_else(WorkerId::now_v7);
             let worker_id_text = worker_id.to_string();
             let now = chrono::Utc::now().to_rfc3339();
             let singleton_generation = match singleton_owner.as_ref() {
@@ -3197,13 +3229,15 @@ impl SqliteWorkspaceStore {
                     ));
                 }
             }
-            tx.commit()?;
-            Ok(WorkerCreateReservation {
+            let reservation = WorkerCreateReservation {
                 worker_id,
                 create_fingerprint,
                 memory_settings: snapshot,
                 singleton,
-            })
+            };
+            complete(&tx, &reservation)?;
+            tx.commit()?;
+            Ok(reservation)
         })
     }
 

crates/workspace-server/src/test_support.rs

@@ -0,0 +1,688 @@
+//! Offline Backend fixture preparation, available only with `test-support`.
+//!
+//! Use a new isolated database and the existing account, Workspace signing-identity /
+//! `ControlPlaneStore::create_workspace_bootstrap`, and Runtime binding APIs first.
+//! Generate fresh account credentials and signing keys for each environment; no fixture
+//! credentials are accepted here. Install Profile configuration through the existing
+//! config-source mechanisms before starting the product processes.
+//!
+//! This module seeds Backend ownership only. It neither writes Runtime storage nor
+//! launches a Worker, fabricates Runtime observations, changes authorization, or sends
+//! initial input. The caller must separately install the matching stopped Runtime
+//! fixture. Build product binaries in a separate Cargo invocation **without** this
+//! feature (Cargo features unify within an invocation).
+
+use super::*;
+use crate::hosts::{WorkerSpawnIntent, WorkerSpawnRequest, worker_spawn_create_fingerprint};
+use worker_runtime::catalog::ProfileSelector;
+
+/// One ordinary, attachment-free Worker from an externally validated fixture.
+///
+/// IDs must match the Runtime fixture. `request` is the unresolved canonical create
+/// intent, not a Runtime launch request containing credentials or execution grants.
+/// The Profile selector must already be installed by the harness. `initial_submit`
+/// must be empty: saved Session contents belong to the Runtime fixture, not new input.
+/// Singleton/service, Ticket-assigned and attachment-bearing Workers are deliberately
+/// unsupported.
+pub struct FixedWorkerSeed<'a> {
+    pub workspace_id: &'a str,
+    pub runtime_id: &'a str,
+    pub worker_id: WorkerId,
+    pub allocation_key: &'a str,
+    pub request: &'a WorkerSpawnRequest,
+}
+
+/// Insert a fixed Worker catalog entry and completed creation reservation atomically.
+///
+/// Uses the normal reservation validators, authoritative Memory snapshot, fingerprint
+/// binding, resource-key allocator and immediate transaction. Even an exact repeat is
+/// rejected: this is not an upsert/import/recovery API. Existing allocations, Worker
+/// identities and residual lifecycle/attachment records are never adopted or repaired.
+/// Returns the creation binding needed to check the corresponding Runtime fixture.
+pub fn seed_fixed_worker(
+    store: &SqliteWorkspaceStore,
+    seed: FixedWorkerSeed<'_>,
+) -> Result<WorkerCreateReservation> {
+    validate_identifier("workspace_id", seed.workspace_id)?;
+    validate_identifier("runtime_id", seed.runtime_id)?;
+    let request = seed.request;
+    if !request.initial_submit.is_empty() {
+        return Err(Error::InvalidInput(
+            "fixture Worker cannot contain initial_submit; offline seeding accepts no new input"
+                .into(),
+        ));
+    }
+    if !matches!(request.intent, WorkerSpawnIntent::WorkspaceCoding)
+        || request.singleton_key.is_some()
+        || request.ticket_assignment.is_some()
+        || request.resolved_subjektiv_attached
+        || !request.workdir_attachment_requests.is_empty()
+        || !request.resolved_workdir_attachment_requests.is_empty()
+        || !request.resolved_workdir_attachments.is_empty()
+        || request.resolved_config_bundle.is_some()
+        || request.resolved_workspace_api.is_some()
+        || request.resolved_memory_settings.is_some()
+        || request.resolved_worker_observation_enabled
+        || !request.resolved_worker_observation_grants.is_empty()
+        || request.resolved_control_operation.is_some()
+        || request.resolved_execution_grant.is_some()
+    {
+        return Err(Error::InvalidInput(
+            "fixture Worker must be ordinary unresolved create intent without singleton, assignments, attachments or grants".into(),
+        ));
+    }
+    // Builtin selectors already carry their canonical `builtin:` prefix.
+    let profile = match &request.profile {
+        ProfileSelector::Builtin(name) | ProfileSelector::Named(name) => name,
+    };
+    validate_non_empty("fixture Profile selector", profile)?;
+    let worker_id = seed.worker_id.to_string();
+    let display_name = request
+        .requested_worker_name
+        .as_deref()
+        .unwrap_or(&worker_id);
+    validate_non_empty("fixture Worker display name", display_name)?;
+    let fingerprint = worker_spawn_create_fingerprint(request).map_err(Error::InvalidInput)?;
+    let memory = store.get_workspace_memory_settings(seed.workspace_id)?;
+    store.reserve_worker_create_with(
+        seed.workspace_id,
+        seed.runtime_id,
+        seed.allocation_key,
+        &fingerprint,
+        None,
+        &memory,
+        |tx| {
+            require_seed_target(tx, &seed)?;
+            Ok(Some(seed.worker_id))
+        },
+        |tx, reservation| {
+            let now = chrono::Utc::now().to_rfc3339();
+            tx.execute(
+                "INSERT INTO worker_registry (
+                    workspace_id, runtime_id, worker_id, display_name, profile,
+                    retention_state, created_at, updated_at
+                 ) VALUES (?1, ?2, ?3, ?4, ?5, 'normal', ?6, ?6)",
+                params![seed.workspace_id, seed.runtime_id, worker_id, display_name, profile, now],
+            )?;
+            let changed = tx.execute(
+                "UPDATE worker_create_reservations SET state = 'created', updated_at = ?4
+                 WHERE workspace_id = ?1 AND runtime_id = ?2 AND worker_id = ?3 AND state = 'reserved'",
+                params![seed.workspace_id, seed.runtime_id, reservation.worker_id.to_string(), now],
+            )?;
+            if changed != 1 {
+                return Err(Error::Store("fixture creation reservation disappeared".into()));
+            }
+            Ok(())
+        },
+    )
+}
+
+fn require_seed_target(tx: &rusqlite::Transaction<'_>, seed: &FixedWorkerSeed<'_>) -> Result<()> {
+    let active: bool = tx.query_row(
+        "SELECT EXISTS(SELECT 1 FROM workspaces w
+         JOIN workspace_runtime_bindings r ON r.workspace_id = w.workspace_id
+         JOIN workspace_signing_identities s ON s.workspace_id = w.workspace_id
+         WHERE w.workspace_id = ?1 AND w.state = 'active'
+           AND r.runtime_id = ?2 AND r.revoked_at IS NULL
+           AND r.authentication_mode = 'workspace_identity' AND s.state = 'active')",
+        params![seed.workspace_id, seed.runtime_id],
+        |row| row.get(0),
+    )?;
+    if !active {
+        return Err(Error::InvalidInput(
+            "fixture requires an active bootstrapped Workspace and authenticated Runtime binding"
+                .into(),
+        ));
+    }
+    let worker_id = seed.worker_id.to_string();
+    let allocation_exists: bool = tx.query_row(
+        "SELECT EXISTS(SELECT 1 FROM worker_create_reservations
+         WHERE workspace_id = ?1 AND allocation_key = ?2)",
+        params![seed.workspace_id, seed.allocation_key],
+        |row| row.get(0),
+    )?;
+    if allocation_exists {
+        return Err(Error::InvalidInput(
+            "fixture Worker allocation already exists".into(),
+        ));
+    }
+    // Check by WorkerId, not just the proposed Runtime tuple: orphaned and terminal
+    // ownership is still a collision. These tables are owned by this store schema.
+    for table in [
+        "worker_create_reservations",
+        "worker_registry",
+        "worker_registry_observations",
+        "worker_registry_projection_removals",
+        "worker_removal_operations",
+        "worker_singleton_owners",
+        "worker_workdir_links",
+        "worker_restore_intents",
+        "worker_tombstones",
+        "worker_orphan_diagnostics",
+        "worker_registry_projection_diagnostics",
+        "worker_session_archives",
+        "worker_diagnostics_archives",
+        "ticket_assignment_worker_tombstones",
+    ] {
+        let exists: bool = tx.query_row(
+            &format!("SELECT EXISTS(SELECT 1 FROM {table} WHERE worker_id = ?1)"),
+            [&worker_id],
+            |row| row.get(0),
+        )?;
+        if exists {
+            return Err(Error::InvalidInput(format!(
+                "fixture Worker identity already exists in {table}"
+            )));
+        }
+    }
+    // A resource key without a catalog/reservation must not be silently reused.
+    let mut statement = tx.prepare(
+        "SELECT resource_id FROM workspace_resource_keys WHERE resource_kind = 'worker'",
+    )?;
+    let resources = statement.query_map([], |row| row.get::<_, String>(0))?;
+    for resource in resources {
+        let resource = resource?;
+        let existing = resource
+            .is_ascii()
+            .then(|| parse_worker_resource_id(&resource))
+            .flatten()
+            .ok_or_else(|| {
+                Error::InvalidInput("unknown fixture Worker resource identity collision".into())
+            })?;
+        if existing.worker_id == worker_id {
+            return Err(Error::InvalidInput(
+                "fixture Worker resource identity already exists".into(),
+            ));
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::hosts::{
+        WorkerSpawnAcceptanceRequirement, WorkerSpawnWorkingDirectoryAttachmentRequest,
+        WorkerSpawnWorkingDirectoryRequest,
+    };
+    use crate::workspace_signing_identity::{
+        InMemoryWorkspaceSigningMaterialStore, WorkspaceSigningIdentityService,
+    };
+
+    const WORKSPACE: &str = "0197a949-4b6b-7f2a-9d9a-1f87e3a4c501";
+    const RUNTIME: &str = "fixture-runtime";
+    const WORKER: &str = "0197a949-4b6b-7f2a-9d9a-1f87e3a4c502";
+
+    // Exercises the same public bootstrap APIs the offline harness uses. All account
+    // and key identities are newly generated, even across stores with fixed fixture IDs.
+    fn bootstrap() -> SqliteWorkspaceStore {
+        let store = SqliteWorkspaceStore::in_memory_for_test().unwrap();
+        let account = AccountRecord {
+            account_id: format!("fixture-account-{}", Uuid::now_v7()),
+            kind: "user".into(),
+            handle: format!("fixture-{}", Uuid::now_v7()),
+            display_name: "Fixture owner".into(),
+            created_at: "2026-10-24T00:00:00Z".into(),
+            updated_at: "2026-10-24T00:00:00Z".into(),
+        };
+        store.upsert_account(&account).unwrap();
+        let identities = WorkspaceSigningIdentityService::new(
+            Arc::new(store.clone()),
+            Arc::new(InMemoryWorkspaceSigningMaterialStore::default()),
+        );
+        let (identity, operation) = identities
+            .prepare_workspace_creation(
+                "fixture-bootstrap",
+                "fixture-bootstrap",
+                WORKSPACE,
+                &account.account_id,
+            )
+            .unwrap();
+        store
+            .create_workspace_bootstrap(
+                &WorkspaceBootstrapRecord {
+                    operation_key: "fixture-bootstrap".into(),
+                    request_fingerprint: "fixture-bootstrap".into(),
+                    workspace: WorkspaceRecord {
+                        workspace_id: WORKSPACE.into(),
+                        owner_account_id: account.account_id,
+                        display_name: "Fixture".into(),
+                        state: "active".into(),
+                        created_at: account.created_at,
+                        updated_at: account.updated_at,
+                    },
+                    repository: None,
+                },
+                &identity,
+                &operation,
+            )
+            .unwrap();
+        let runtime_key = worker_runtime::auth::RuntimeIdentityMaterial::generate(RUNTIME).unwrap();
+        store
+            .upsert_workspace_runtime_binding(
+                WorkspaceRuntimeBinding {
+                    workspace_id: WORKSPACE.into(),
+                    runtime_id: RUNTIME.into(),
+                    display_name: "Fixture Runtime".into(),
+                    base_url: "http://127.0.0.1:12345".into(),
+                    public_key: runtime_key.public_key,
+                    public_key_fingerprint: String::new(),
+                    binding_id: String::new(),
+                    authentication_mode: WorkspaceRuntimeAuthenticationMode::WorkspaceIdentity,
+                    created_at: "2026-10-24T00:00:00Z".into(),
+                    updated_at: "2026-10-24T00:00:00Z".into(),
+                    revoked_at: None,
+                },
+                false,
+            )
+            .unwrap();
+        store
+    }
+
+    fn request() -> WorkerSpawnRequest {
+        WorkerSpawnRequest {
+            intent: WorkerSpawnIntent::WorkspaceCoding,
+            requested_worker_name: Some("Fixture Worker".into()),
+            singleton_key: None,
+            acceptance: WorkerSpawnAcceptanceRequirement::SocketReady,
+            profile: ProfileSelector::Builtin("builtin:coder".into()),
+            ticket_assignment: None,
+            initial_submit: Vec::new(),
+            workdir_attachment_requests: Vec::new(),
+            resolved_workdir_attachment_requests: Vec::new(),
+            resolved_workdir_attachments: Vec::new(),
+            resolved_config_bundle: None,
+            resolved_workspace_api: None,
+            resolved_memory_settings: None,
+            resolved_subjektiv_attached: false,
+            resolved_worker_observation_enabled: false,
+            resolved_worker_observation_grants: Vec::new(),
+            resolved_control_operation: None,
+            resolved_execution_grant: None,
+        }
+    }
+
+    fn seed(request: &WorkerSpawnRequest) -> FixedWorkerSeed<'_> {
+        FixedWorkerSeed {
+            workspace_id: WORKSPACE,
+            runtime_id: RUNTIME,
+            worker_id: WORKER.parse().unwrap(),
+            allocation_key: "fixture-allocation",
+            request,
+        }
+    }
+
+    fn worker() -> RuntimeWorkerRef {
+        RuntimeWorkerRef::new(RUNTIME, WORKER)
+    }
+
+    #[test]
+    fn fixed_ids_are_isolated_and_seed_has_no_runtime_observation() {
+        let a = bootstrap();
+        let b = bootstrap();
+        let request = request();
+        let first = seed_fixed_worker(&a, seed(&request)).unwrap();
+        let second = seed_fixed_worker(&b, seed(&request)).unwrap();
+        assert_eq!(first, second);
+        assert_eq!(first.worker_id.to_string(), WORKER);
+        assert_eq!(first.memory_settings.workspace_id, WORKSPACE);
+        assert_eq!(first.memory_settings.language, "English");
+        assert!(first.singleton.is_none());
+        assert_ne!(
+            a.get_workspace_signing_identity(WORKSPACE)
+                .unwrap()
+                .unwrap()
+                .public_key,
+            b.get_workspace_signing_identity(WORKSPACE)
+                .unwrap()
+                .unwrap()
+                .public_key
+        );
+        assert_ne!(
+            a.get_workspace_runtime_binding(WORKSPACE, RUNTIME)
+                .unwrap()
+                .unwrap()
+                .public_key,
+            b.get_workspace_runtime_binding(WORKSPACE, RUNTIME)
+                .unwrap()
+                .unwrap()
+                .public_key
+        );
+        for store in [&a, &b] {
+            assert_eq!(
+                store
+                    .worker_resource_key(WORKSPACE, &worker())
+                    .unwrap()
+                    .as_deref(),
+                Some("W-1")
+            );
+            assert_eq!(
+                store
+                    .resolve_worker_resource_reference(WORKSPACE, "W-1")
+                    .unwrap(),
+                Some(worker())
+            );
+            assert!(
+                !store
+                    .has_active_worker_create_reservation(WORKSPACE, &worker())
+                    .unwrap()
+            );
+            let registry = store
+                .get_worker_registry(WORKSPACE, &worker())
+                .unwrap()
+                .unwrap();
+            assert_eq!(registry.profile.as_deref(), Some("builtin:coder"));
+            assert_eq!(registry.display_name, "Fixture Worker");
+            assert_eq!(registry.retention_state, "normal");
+            let projection = store
+                .with_conn(|conn| read_worker_registry_projection(conn, WORKSPACE, &worker()))
+                .unwrap()
+                .unwrap();
+            assert!(projection.observation.is_none());
+            assert!(projection.job.is_none());
+        }
+    }
+
+    #[test]
+    fn initial_submit_is_refused_without_registration_or_allocation() {
+        let store = bootstrap();
+        let mut request = request();
+        // Even an empty text segment is new input; the vector itself must be empty.
+        for text in ["new input must not be seeded", ""] {
+            request.initial_submit = vec![protocol::Segment::text(text)];
+            let error = seed_fixed_worker(&store, seed(&request)).unwrap_err();
+            assert!(
+                matches!(error, Error::InvalidInput(ref message) if message.contains("initial_submit"))
+            );
+            assert!(
+                store
+                    .get_worker_registry(WORKSPACE, &worker())
+                    .unwrap()
+                    .is_none()
+            );
+            assert!(
+                store
+                    .worker_resource_key(WORKSPACE, &worker())
+                    .unwrap()
+                    .is_none()
+            );
+            assert!(
+                !store
+                    .has_active_worker_create_reservation(WORKSPACE, &worker())
+                    .unwrap()
+            );
+        }
+        // The rejected attempts must not consume the allocation or resource counter.
+        request.initial_submit.clear();
+        seed_fixed_worker(&store, seed(&request)).unwrap();
+        assert_eq!(
+            store
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .as_deref(),
+            Some("W-1")
+        );
+    }
+
+    #[test]
+    fn exact_replay_changed_input_and_cross_runtime_identity_are_refused() {
+        let store = bootstrap();
+        let mut request = request();
+        seed_fixed_worker(&store, seed(&request)).unwrap();
+        let before = store.get_worker_registry(WORKSPACE, &worker()).unwrap();
+        assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        request.requested_worker_name = Some("Must not overwrite".into());
+        assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        let mut new_allocation = seed(&request);
+        new_allocation.allocation_key = "another-allocation";
+        assert!(seed_fixed_worker(&store, new_allocation).is_err());
+        let mut same_allocation = seed(&request);
+        same_allocation.worker_id = WorkerId::now_v7();
+        assert!(seed_fixed_worker(&store, same_allocation).is_err());
+        assert_eq!(
+            store.get_worker_registry(WORKSPACE, &worker()).unwrap(),
+            before
+        );
+
+        let other = bootstrap();
+        let mut record = before.unwrap();
+        record.worker.runtime_id = "different-runtime".into();
+        other.upsert_worker_registry(&record).unwrap();
+        assert!(seed_fixed_worker(&other, seed(&request)).is_err());
+        assert!(
+            other
+                .get_worker_registry(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn reserved_and_orphaned_resource_identities_are_not_adopted() {
+        let store = bootstrap();
+        let request = request();
+        let memory = store.get_workspace_memory_settings(WORKSPACE).unwrap();
+        let ordinary = store
+            .reserve_worker_create(WORKSPACE, RUNTIME, "ordinary", "fingerprint", None, &memory)
+            .unwrap();
+        let mut collision = seed(&request);
+        collision.worker_id = ordinary.worker_id;
+        assert!(seed_fixed_worker(&store, collision).is_err());
+        // Store-level fault setup only: the external harness never writes SQL.
+        store
+            .with_conn(|conn| {
+                allocate_resource_key(
+                    conn,
+                    WORKSPACE,
+                    WorkspaceResourceKind::Worker,
+                    &worker_resource_id(&worker()),
+                    "now",
+                )?;
+                Ok(())
+            })
+            .unwrap();
+        assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        assert!(
+            store
+                .get_worker_registry(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn unknown_resource_identity_and_retained_removal_fence_are_refused() {
+        let request = request();
+        let malformed = bootstrap();
+        malformed
+            .with_conn(|conn| {
+                allocate_resource_key(
+                    conn,
+                    WORKSPACE,
+                    WorkspaceResourceKind::Worker,
+                    "unknown-identity",
+                    "now",
+                )?;
+                Ok(())
+            })
+            .unwrap();
+        assert!(matches!(
+            seed_fixed_worker(&malformed, seed(&request)),
+            Err(Error::InvalidInput(_))
+        ));
+        assert!(
+            malformed
+                .get_worker_registry(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+
+        let removed = bootstrap();
+        removed.with_conn(|conn| {
+            conn.execute(
+                "INSERT INTO worker_registry_projection_removals(workspace_id, runtime_id, worker_id) VALUES (?1, ?2, ?3)",
+                params![WORKSPACE, RUNTIME, WORKER],
+            )?;
+            Ok(())
+        }).unwrap();
+        assert!(matches!(
+            seed_fixed_worker(&removed, seed(&request)),
+            Err(Error::InvalidInput(_))
+        ));
+        assert!(
+            removed
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn unknown_workspace_runtime_and_revoked_binding_are_refused() {
+        let store = bootstrap();
+        let request = request();
+        let mut unknown = seed(&request);
+        unknown.workspace_id = "unknown";
+        assert!(seed_fixed_worker(&store, unknown).is_err());
+        let mut unknown = seed(&request);
+        unknown.runtime_id = "unknown";
+        assert!(seed_fixed_worker(&store, unknown).is_err());
+        store
+            .revoke_workspace_runtime_binding(WORKSPACE, RUNTIME, "2026-10-24T01:00:00Z")
+            .unwrap();
+        assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        assert!(
+            store
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn unsupported_intent_is_refused_without_partial_registration() {
+        let store = bootstrap();
+        let base = request();
+        let mut requests = Vec::new();
+        let mut request = base.clone();
+        request.singleton_key = Some("fixture-singleton".into());
+        requests.push(request);
+        let mut request = base.clone();
+        request.intent = WorkerSpawnIntent::WorkspaceOrchestrator;
+        requests.push(request);
+        let mut request = base.clone();
+        request.resolved_subjektiv_attached = true;
+        requests.push(request);
+        let mut request = base.clone();
+        request
+            .workdir_attachment_requests
+            .push(WorkerSpawnWorkingDirectoryAttachmentRequest {
+                alias: workdir::WorkdirAttachmentAlias::new("repo").unwrap(),
+                working_directory: WorkerSpawnWorkingDirectoryRequest {
+                    repository_key: "repo".into(),
+                    selector: None,
+                },
+            });
+        requests.push(request);
+        for request in requests {
+            assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        }
+        assert!(
+            store
+                .get_worker_registry(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            store
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+        assert_eq!(
+            seed_fixed_worker(&store, seed(&base))
+                .unwrap()
+                .worker_id
+                .to_string(),
+            WORKER
+        );
+    }
+
+    #[test]
+    fn late_catalog_failure_rolls_back_reservation_and_resource_allocation() {
+        let store = bootstrap();
+        let request = request();
+        store
+            .with_conn(|conn| {
+                conn.execute_batch(
+                    "CREATE TEMP TRIGGER fail_fixture_catalog BEFORE INSERT ON worker_registry
+                                BEGIN SELECT RAISE(ABORT, 'injected catalog failure'); END",
+                )?;
+                Ok(())
+            })
+            .unwrap();
+        assert!(seed_fixed_worker(&store, seed(&request)).is_err());
+        assert!(
+            !store
+                .has_active_worker_create_reservation(WORKSPACE, &worker())
+                .unwrap()
+        );
+        assert!(
+            store
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .is_none()
+        );
+        store
+            .with_conn(|conn| {
+                conn.execute_batch("DROP TRIGGER fail_fixture_catalog")?;
+                Ok(())
+            })
+            .unwrap();
+        seed_fixed_worker(&store, seed(&request)).unwrap();
+        assert_eq!(
+            store
+                .worker_resource_key(WORKSPACE, &worker())
+                .unwrap()
+                .as_deref(),
+            Some("W-1")
+        );
+    }
+
+    #[test]
+    fn ordinary_creation_still_allocates_and_replays_without_catalog_insertion() {
+        let store = bootstrap();
+        let memory = store.get_workspace_memory_settings(WORKSPACE).unwrap();
+        let reserve = |key| {
+            store
+                .reserve_worker_create(WORKSPACE, RUNTIME, key, "fingerprint", None, &memory)
+                .unwrap()
+        };
+        let first = reserve("ordinary-one");
+        assert_eq!(first, reserve("ordinary-one"));
+        let second = reserve("ordinary-two");
+        assert_ne!(first.worker_id, second.worker_id);
+        assert_ne!(first.worker_id.to_string(), WORKER);
+        assert_eq!(
+            first.worker_id.as_uuid().get_version(),
+            Some(uuid::Version::SortRand)
+        );
+        let worker = RuntimeWorkerRef::new(RUNTIME, first.worker_id.to_string());
+        assert!(
+            store
+                .has_active_worker_create_reservation(WORKSPACE, &worker)
+                .unwrap()
+        );
+        assert!(
+            store
+                .get_worker_registry(WORKSPACE, &worker)
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            store
+                .reserve_worker_create(WORKSPACE, RUNTIME, "ordinary-one", "changed", None, &memory)
+                .is_err()
+        );
+    }
+}