fix: recover unpublished Subject creation and preserve auth rejections

crates/runtime-api/src/lib.rs

@@ -1140,8 +1140,36 @@ pub struct SubjectReleaseResponse {
     pub state: SubjectReleaseState,
 }
 
+/// Irreversibly fence an unpublished failed Subject create. Exact retries must retain
+/// all four fields; this is distinct from releasing an existing Subject Controller.
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct SubjectCreateRecoveryRequest {
+    pub workspace_id: String,
+    pub worker_id: WorkerId,
+    pub create_fingerprint: String,
+    pub operation_id: String,
+}
+
+/// Durable admission-fence proof plus freshly confirmed execution absence.
+/// An error (including a lost response) is never evidence of absence.
+#[derive(Clone, Debug, Deserialize, Eq, PartialEq, Serialize)]
+#[serde(deny_unknown_fields)]
+pub struct SubjectCreateRecoveryResponse {
+    pub workspace_id: String,
+    pub worker_id: WorkerId,
+    pub create_fingerprint: String,
+    pub operation_id: String,
+}
+
 #[api(reqwest, axum)]
 pub trait RuntimeApi {
+    #[post("/v1/subject-create-recovery", status = 200, error_status = 400)]
+    async fn recover_subject_create(
+        &self,
+        #[body] request: SubjectCreateRecoveryRequest,
+    ) -> Result<SubjectCreateRecoveryResponse, RuntimeApiError>;
+
     #[post("/v1/subject-release", status = 200, error_status = 400)]
     async fn subject_release(
         &self,
@@ -1675,6 +1703,18 @@ mod tests {
         ) -> Result<SessionMaterialPinResponse, RuntimeApiError> {
             Err(test_error(501))
         }
+        async fn recover_subject_create(
+            &self,
+            request: SubjectCreateRecoveryRequest,
+        ) -> Result<SubjectCreateRecoveryResponse, RuntimeApiError> {
+            Ok(SubjectCreateRecoveryResponse {
+                workspace_id: request.workspace_id,
+                worker_id: request.worker_id,
+                create_fingerprint: request.create_fingerprint,
+                operation_id: request.operation_id,
+            })
+        }
+
         async fn subject_release(
             &self,
             _request: SubjectReleaseRequest,
@@ -1820,6 +1860,21 @@ mod tests {
             .build()
             .unwrap();
 
+        let recovery = SubjectCreateRecoveryRequest {
+            workspace_id: "workspace-a".into(),
+            worker_id: WorkerId::now_v7(),
+            create_fingerprint: "failed-create-fingerprint".into(),
+            operation_id: "recovery-1".into(),
+        };
+        let response = client
+            .recover_subject_create(recovery.clone())
+            .await
+            .unwrap();
+        assert_eq!(response.workspace_id, recovery.workspace_id);
+        assert_eq!(response.worker_id, recovery.worker_id);
+        assert_eq!(response.create_fingerprint, recovery.create_fingerprint);
+        assert_eq!(response.operation_id, recovery.operation_id);
+
         let summary = client.runtime_summary().await.unwrap();
         assert_eq!(summary.runtime.display_name.as_deref(), Some("round-trip"));
         let workers = client

crates/subjektiv/src/journal/binding.rs

@@ -67,6 +67,27 @@ impl SubjektivStore {
         })
     }
 
+    /// Absence proof for failed body creation, including bindings never attached to a Session
+    /// and released lifetimes. The Host must hold its lifecycle fence across this observation
+    /// and recovery; absence alone does not fence a delayed Runtime create.
+    pub fn journal_body_is_unpublished(
+        &self,
+        subject: &str,
+        runtime: &str,
+        worker: &str,
+    ) -> JournalResult<bool> {
+        self.database.try_with_connection(|tx| {
+            Ok(tx.query_row(
+                "SELECT NOT EXISTS(SELECT 1 FROM journal_writer_bindings
+                   WHERE subject_id=?1 AND runtime_id=?2 AND worker_id=?3)
+                 AND NOT EXISTS(SELECT 1 FROM subject_session_attributions
+                   WHERE subject_id=?1 AND runtime_id=?2 AND worker_id=?3)",
+                params![subject, runtime, worker],
+                |row| row.get(0),
+            )?)
+        })
+    }
+
     pub fn journal_worker_was_released(
         &self,
         subject: &str,
@@ -203,3 +224,82 @@ impl SubjektivStore {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn unpublished_body_rejects_orphan_and_released_bindings_and_attribution_only() {
+        let temp = tempfile::tempdir().unwrap();
+        let (_manager, _workspace, store) = crate::tests::open_store(temp.path(), "absence-tests");
+        let subject = store
+            .create_subject(crate::SubjectRole::new("companion").unwrap())
+            .unwrap();
+        let other_subject = store
+            .create_subject(crate::SubjectRole::new("companion").unwrap())
+            .unwrap();
+        assert!(
+            store
+                .journal_body_is_unpublished(&subject.id, "runtime", "worker")
+                .unwrap()
+        );
+        // No Session was attached: an orphan binding is still published body evidence.
+        let binding = store
+            .claim_journal_writer(&subject.id, "runtime", "worker")
+            .unwrap();
+        assert!(
+            !store
+                .journal_body_is_unpublished(&subject.id, "runtime", "worker")
+                .unwrap()
+        );
+        for (subject_id, runtime, worker) in [
+            (other_subject.id.as_str(), "runtime", "worker"),
+            (subject.id.as_str(), "other-runtime", "worker"),
+            (subject.id.as_str(), "runtime", "other-worker"),
+        ] {
+            assert!(
+                store
+                    .journal_body_is_unpublished(subject_id, runtime, worker)
+                    .unwrap()
+            );
+        }
+        store.release_journal_writer(&binding).unwrap();
+        assert!(
+            !store
+                .journal_body_is_unpublished(&subject.id, "runtime", "worker")
+                .unwrap()
+        );
+        store
+            .record_session_attribution(
+                SubjectSessionAttribution::new(
+                    &subject.id,
+                    "runtime",
+                    "attribution-only",
+                    "session",
+                )
+                .unwrap(),
+            )
+            .unwrap();
+        assert!(
+            !store
+                .journal_body_is_unpublished(&subject.id, "runtime", "attribution-only")
+                .unwrap()
+        );
+        assert!(
+            store
+                .journal_body_is_unpublished(&subject.id, "other-runtime", "attribution-only")
+                .unwrap()
+        );
+        assert!(
+            store
+                .journal_body_is_unpublished(&other_subject.id, "runtime", "attribution-only")
+                .unwrap()
+        );
+        assert!(
+            store
+                .journal_body_is_unpublished(&subject.id, "runtime", "never-published")
+                .unwrap()
+        );
+    }
+}

crates/worker-runtime/src/execution.rs

@@ -567,6 +567,15 @@ impl fmt::Debug for WorkerProtocolTransport {
 pub trait WorkerExecutionBackend: Send + Sync + 'static {
     fn backend_id(&self) -> &str;
 
+    /// Confirm actual absence under the backend Worker operation lock. Runtime has
+    /// already durably fenced admission; unsupported or uncertain backends must fail.
+    fn recover_subject_create(
+        &self,
+        _request: &runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<(), String> {
+        Err("execution backend cannot confirm Subject create execution absence".into())
+    }
+
     /// Must reserve the actual Controller, not infer Idle from a catalog snapshot.
     fn subject_release(
         &self,
@@ -839,6 +848,13 @@ impl WorkerExecutionBackendRef {
         self.backend.fetch_workspace_config(request)
     }
 
+    pub(crate) fn recover_subject_create(
+        &self,
+        request: &runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<(), String> {
+        self.backend.recover_subject_create(request)
+    }
+
     pub(crate) fn subject_release(
         &self,
         request: &runtime_api::SubjectReleaseRequest,

crates/worker-runtime/src/fs_store.rs

@@ -39,9 +39,10 @@ use std::path::{Path, PathBuf};
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU8, AtomicU64, Ordering};
 
-const SCHEMA_VERSION: u32 = 11;
-const EXECUTION_SCHEMA_VERSION: u32 = 10;
+const SCHEMA_VERSION: u32 = 12;
 // Frozen source schemas: do not move their meaning when adding a migration.
+const SOURCE_SCHEMA_VERSION: u32 = 11;
+const EXECUTION_SCHEMA_VERSION: u32 = 10;
 const AGGREGATE_SCHEMA_VERSION: u32 = 9;
 const PREVIOUS_SCHEMA_VERSION: u32 = 8;
 const LEGACY_SCHEMA_VERSION: u32 = 7;
@@ -480,6 +481,8 @@ pub(crate) struct OpenedFsRuntimeStore {
 
 #[derive(Clone, Debug)]
 pub(crate) struct PersistedRuntimeState {
+    pub(crate) subject_create_recovery_fences:
+        BTreeMap<WorkerId, crate::runtime::SubjectCreateRecoveryFence>,
     pub(crate) subject_releases: Vec<(
         runtime_api::SubjectReleaseRequest,
         runtime_api::SubjectReleaseState,
@@ -641,8 +644,8 @@ fn plan_runtime_store_migration(
             format!("Runtime store schema version {schema_version} is out of range"),
         )
     })?;
-    let staging = migration_sibling(root, "schema-v11-staging")?;
-    let backup = migration_sibling(root, "pre-schema-v11-backup")?;
+    let staging = migration_sibling(root, "schema-v12-staging")?;
+    let backup = migration_sibling(root, "pre-schema-v12-backup")?;
     if staging.exists() || backup.exists() {
         return Err(runtime_store_corrupt(
             root,
@@ -676,11 +679,12 @@ fn plan_runtime_store_migration(
             | PREVIOUS_SCHEMA_VERSION
             | AGGREGATE_SCHEMA_VERSION
             | EXECUTION_SCHEMA_VERSION
+            | SOURCE_SCHEMA_VERSION
     ) {
         return Err(runtime_store_corrupt(
             &runtime_path,
             format!(
-                "unsupported Runtime store schema version {schema_version}; expected {OLDEST_SCHEMA_VERSION}, {LEGACY_SCHEMA_VERSION}, {PREVIOUS_SCHEMA_VERSION}, {AGGREGATE_SCHEMA_VERSION}, {EXECUTION_SCHEMA_VERSION}, or {SCHEMA_VERSION}"
+                "unsupported Runtime store schema version {schema_version}; expected {OLDEST_SCHEMA_VERSION}, {LEGACY_SCHEMA_VERSION}, {PREVIOUS_SCHEMA_VERSION}, {AGGREGATE_SCHEMA_VERSION}, {EXECUTION_SCHEMA_VERSION}, {SOURCE_SCHEMA_VERSION}, or {SCHEMA_VERSION}"
             ),
         ));
     }
@@ -1010,11 +1014,16 @@ fn migrate_worker_record(
     mapping: Option<&LegacyWorkerIdentityMapping>,
     identity_path: &Path,
 ) -> Result<WorkerAggregateRecord, RuntimeError> {
-    if source_schema_version == EXECUTION_SCHEMA_VERSION {
+    if matches!(
+        source_schema_version,
+        EXECUTION_SCHEMA_VERSION | SOURCE_SCHEMA_VERSION
+    ) {
+        // Schemas 10 and 11 already use the current Worker aggregate layout.
+        // Only advance its version; never regenerate Restore admission identity.
         if identity_document
             .get("schema_version")
             .and_then(serde_json::Value::as_u64)
-            != Some(u64::from(EXECUTION_SCHEMA_VERSION))
+            != Some(u64::from(source_schema_version))
         {
             return Err(runtime_store_corrupt(
                 identity_path,
@@ -1405,11 +1414,18 @@ fn migrate_runtime_document(
         "schema_version".to_string(),
         serde_json::Value::from(SCHEMA_VERSION),
     );
-    object.remove("workers");
-    // This legacy cache is not Runtime lifecycle or restore authority. Bundles
-    // are reacquired from Workspace content authority, never renumbered/rehash-pinned here.
-    object.remove("config_bundles");
-    object.remove("next_worker_sequence");
+    // Intermediate schema-11 writers may already have saved irreversible fences.
+    // Default only a missing field; malformed or existing fences must not be erased.
+    object
+        .entry("subject_create_recovery_fences")
+        .or_insert_with(|| serde_json::json!({}));
+    if source_schema_version != SOURCE_SCHEMA_VERSION {
+        object.remove("workers");
+        // This legacy cache is not Runtime lifecycle or restore authority. Bundles
+        // are reacquired from Workspace content authority, never renumbered/rehash-pinned here.
+        object.remove("config_bundles");
+        object.remove("next_worker_sequence");
+    }
     let snapshot: RuntimeSnapshot = serde_json::from_value(document.clone()).map_err(|error| {
         runtime_store_corrupt(
             runtime_path,
@@ -1523,14 +1539,16 @@ fn recover_runtime_store_migration(root: &Path) -> Result<(), RuntimeError> {
         "pre-schema-v9-backup",
         "schema-v10-staging",
         "pre-schema-v10-backup",
+        "schema-v11-staging",
+        "pre-schema-v11-backup",
     ] {
         let earlier = migration_sibling(root, suffix)?;
         if earlier.exists() {
-            return Err(runtime_store_corrupt(root, "unfinished earlier Runtime migration requires explicit recovery before schema-v11 startup".into()));
+            return Err(runtime_store_corrupt(root, "unfinished earlier Runtime migration requires explicit recovery before schema-v12 startup".into()));
         }
     }
-    let staging = migration_sibling(root, "schema-v11-staging")?;
-    let backup = migration_sibling(root, "pre-schema-v11-backup")?;
+    let staging = migration_sibling(root, "schema-v12-staging")?;
+    let backup = migration_sibling(root, "pre-schema-v12-backup")?;
     let root_exists = migration_directory_exists(root)?;
     let staging_exists = migration_directory_exists(&staging)?;
     let backup_exists = migration_directory_exists(&backup)?;
@@ -1732,8 +1750,8 @@ fn migrate_runtime_store(
     if !plan.migration_required {
         return Ok(plan);
     }
-    let staging = migration_sibling(root, "schema-v11-staging")?;
-    let backup = migration_sibling(root, "pre-schema-v11-backup")?;
+    let staging = migration_sibling(root, "schema-v12-staging")?;
+    let backup = migration_sibling(root, "pre-schema-v12-backup")?;
     if staging.exists() || backup.exists() {
         return Err(runtime_store_corrupt(
             root,
@@ -1908,7 +1926,10 @@ fn migrate_runtime_store_in_place(
                 ));
             }
         }
-        if let Some(metadata) = metadata {
+        // Schema 11 metadata was validated above but needs no conversion.
+        if let Some(metadata) = metadata
+            && plan.current_schema_version != SOURCE_SCHEMA_VERSION
+        {
             atomic_write_json(
                 &migrated_dir.join(WORKER_METADATA_FILE),
                 &metadata,
@@ -1927,7 +1948,11 @@ fn migrate_runtime_store_in_place(
             ),
         ));
     }
-    crate::retention::migrate_staged_sessions(root)?;
+    // Schema 11 already migrated Sessions; the recovery-fence bump must not
+    // rewrite retained history or archive attestations.
+    if plan.current_schema_version != SOURCE_SCHEMA_VERSION {
+        crate::retention::migrate_staged_sessions(root)?;
+    }
     let (document, _) = migrate_runtime_document(
         document,
         plan.current_schema_version,
@@ -1944,6 +1969,8 @@ fn migrate_runtime_store_in_place(
 
 #[derive(Clone, Debug, Serialize, Deserialize)]
 struct RuntimeSnapshot {
+    #[serde(default)]
+    subject_create_recovery_fences: BTreeMap<WorkerId, crate::runtime::SubjectCreateRecoveryFence>,
     #[serde(default)]
     subject_releases: Vec<(
         runtime_api::SubjectReleaseRequest,
@@ -1980,6 +2007,7 @@ fn record_worker_load_diagnostic(
 impl RuntimeSnapshot {
     fn from_persisted(state: &PersistedRuntimeState) -> Self {
         Self {
+            subject_create_recovery_fences: state.subject_create_recovery_fences.clone(),
             subject_releases: state.subject_releases.clone(),
             schema_version: SCHEMA_VERSION,
             display_name: state.display_name.clone(),
@@ -2003,6 +2031,33 @@ impl RuntimeSnapshot {
                 ),
             });
         }
+        for (worker_id, fence) in &self.subject_create_recovery_fences {
+            let request = &fence.request;
+            let response_matches = fence.response.as_ref().is_none_or(|response| {
+                response.worker_id == request.worker_id
+                    && response.workspace_id == request.workspace_id
+                    && response.create_fingerprint == request.create_fingerprint
+                    && response.operation_id == request.operation_id
+            });
+            if *worker_id != request.worker_id
+                || request.workspace_id.trim().is_empty()
+                || request.create_fingerprint.trim().is_empty()
+                || request.create_fingerprint.len() > 512
+                || request.operation_id.trim().is_empty()
+                || request.operation_id.len() > 512
+                || self
+                    .workspace_owners
+                    .get(&request.workspace_id)
+                    .is_none_or(|owner| owner.trim().is_empty())
+                || !response_matches
+            {
+                return Err(RuntimeError::StoreCorrupt {
+                    operation: "read Subject create recovery fence",
+                    path: path.to_path_buf(),
+                    message: "invalid recovery identity, receipt or Workspace owner".into(),
+                });
+            }
+        }
         if self.backend != RuntimeBackendKind::FsStore {
             return Err(RuntimeError::StoreCorrupt {
                 operation: "read runtime snapshot",
@@ -2018,6 +2073,7 @@ impl RuntimeSnapshot {
         workers: BTreeMap<WorkerId, PersistedWorkerRecord>,
     ) -> PersistedRuntimeState {
         PersistedRuntimeState {
+            subject_create_recovery_fences: self.subject_create_recovery_fences,
             subject_releases: self.subject_releases,
             display_name: self.display_name,
             status: self.status,
@@ -2708,6 +2764,199 @@ mod tests {
         document["schema_version"].as_u64().unwrap()
     }
 
+    #[test]
+    fn schema_v11_migration_preserves_identity_history_and_recovery_fences_on_reopen() {
+        // Cover the released schema (no field) and intermediate writers with
+        // either a pending fence or a completed recovery receipt.
+        for fence_state in ["absent", "pending", "completed"] {
+            let temp = tempfile::tempdir().unwrap();
+            let root = temp.path().join("runtime");
+            let worker_id = WorkerId::now_v7();
+            let fenced_worker_id = WorkerId::now_v7();
+            let worker_dir = root.join(WORKERS_DIR).join(worker_id.to_string());
+            fs::create_dir_all(worker_dir.join("session")).unwrap();
+            let runtime_path = root.join(RUNTIME_FILE);
+            let worker_path = worker_dir.join(WORKER_FILE);
+            let metadata_path = worker_dir.join(WORKER_METADATA_FILE);
+            let history_path = worker_dir.join("session/history.json");
+            let history = b"[{\"role\":\"user\",\"content\":\"retained history\"}]\n";
+            fs::write(&history_path, history).unwrap();
+            let metadata = serde_json::to_vec(&serde_json::json!({
+                "worker_name": runtime_worker_name(worker_id),
+                "workspace_id": "workspace-a",
+                "retained_extension": {"history": "must not be rewritten"}
+            }))
+            .unwrap();
+            fs::write(&metadata_path, &metadata).unwrap();
+
+            let mut runtime = serde_json::json!({
+                "schema_version": 11,
+                "display_name": "schema-11 Runtime",
+                "backend": "fs_store",
+                "status": "running",
+                "next_diagnostic_id": 2,
+                "config_bundles": {},
+                "workspace_owners": {"workspace-a": "owner-a"},
+                "subject_releases": [],
+                "diagnostics": [{
+                    "id": 1, "worker_ref": {"worker_id": worker_id},
+                    "severity": "warning", "code": "retained_history",
+                    "message": "preserve diagnostic identity"
+                }]
+            });
+            let recovery_identity = serde_json::json!({
+                "workspace_id": "workspace-a", "worker_id": fenced_worker_id,
+                "create_fingerprint": "failed-create", "operation_id": "recover-a"
+            });
+            if fence_state != "absent" {
+                runtime["subject_create_recovery_fences"] = serde_json::json!({
+                    fenced_worker_id.to_string(): {
+                        "request": recovery_identity,
+                        "response": if fence_state == "completed" {
+                            recovery_identity.clone()
+                        } else {
+                            serde_json::Value::Null
+                        }
+                    }
+                });
+            }
+            atomic_write_json(&runtime_path, &runtime, "write schema-11 fixture").unwrap();
+            let runtime_before = fs::read(&runtime_path).unwrap();
+
+            let mut request = schema_v7_worker_document(worker_id)["request"].clone();
+            request["workspace_api"] = serde_json::json!({
+                "workspace_id": "workspace-a", "base_url": "https://workspace.invalid"
+            });
+            // Serialize defaults as a schema-11 writer would, without running
+            // a migration or generating a fresh Restore guard for the fixture.
+            let request: CreateWorkerRequest = serde_json::from_value(request).unwrap();
+            let worker: WorkerAggregateRecord = serde_json::from_value(serde_json::json!({
+                "schema_version": 11,
+                "restore_guard": {
+                    "observation_token": "retained-restore-token",
+                    "active_request_id": null, "owners": {}
+                },
+                "worker_ref": {"worker_id": worker_id}, "worker_id": worker_id,
+                "profile": request.profile,
+                "display_name": null,
+                "profile_source": request.profile_source.reference(),
+                "config_bundle": null,
+                "created_at_ms": 1234,
+                "status": "running",
+                "last_finished_submission_request_id": "finished-submission",
+                "workspace_id": "workspace-a",
+                "execution_state": {"state": "available", "execution": {
+                    "request": request, "binding": {}, "restore_intent": "automatic"
+                }}
+            }))
+            .unwrap();
+            let mut expected_worker = serde_json::to_value(&worker).unwrap();
+            atomic_write_json(&worker_path, &worker, "write schema-11 Worker").unwrap();
+            let worker_before = fs::read(&worker_path).unwrap();
+            let options = FsRuntimeStoreOptions::new(&root).with_runtime_id("runtime-test");
+            let plan = FsRuntimeStore::migration_plan(&options).unwrap();
+            assert_eq!(plan.current_schema_version, 11);
+            assert_eq!(plan.target_schema_version, 12);
+            assert!(plan.migration_required);
+            assert_eq!(plan.worker_count, 1);
+            assert_eq!(plan.migrated_worker_aggregate_count, 0);
+            assert!(plan.mappings.is_empty());
+            assert_eq!(fs::read(&runtime_path).unwrap(), runtime_before);
+            assert_eq!(fs::read(&worker_path).unwrap(), worker_before);
+            // The direct reader still rejects a previous schema: migration is
+            // explicit, not permissive deserialization of a newer format.
+            let snapshot: RuntimeSnapshot = serde_json::from_value(runtime.clone()).unwrap();
+            assert!(snapshot.validate(&runtime_path).is_err());
+
+            let opened = FsRuntimeStore::open_or_create(root.clone(), "runtime-test").unwrap();
+            let state = opened.state.as_ref().unwrap();
+            let persisted_worker = state.workers.get(&worker_id).unwrap();
+            expected_worker["schema_version"] = serde_json::json!(12);
+            assert_eq!(
+                serde_json::to_value(WorkerAggregateRecord::from_persisted(persisted_worker))
+                    .unwrap(),
+                expected_worker
+            );
+            runtime["schema_version"] = serde_json::json!(12);
+            if fence_state == "absent" {
+                runtime["subject_create_recovery_fences"] = serde_json::json!({});
+            }
+            assert_eq!(
+                serde_json::to_value(&state.subject_create_recovery_fences).unwrap(),
+                runtime["subject_create_recovery_fences"]
+            );
+            assert_eq!(
+                read_json::<serde_json::Value>(&runtime_path, "read migrated Runtime").unwrap(),
+                runtime
+            );
+            assert_eq!(
+                read_json::<serde_json::Value>(&worker_path, "read migrated Worker").unwrap(),
+                expected_worker
+            );
+            // Exercise the normal writer too: it must retain intermediate fences.
+            opened.store.write_runtime_snapshot(state).unwrap();
+            drop(opened);
+            let runtime_once = fs::read(&runtime_path).unwrap();
+            let worker_once = fs::read(&worker_path).unwrap();
+            let reopened = FsRuntimeStore::open_or_create(root.clone(), "runtime-test").unwrap();
+            assert_eq!(
+                serde_json::to_value(
+                    &reopened
+                        .state
+                        .as_ref()
+                        .unwrap()
+                        .subject_create_recovery_fences
+                )
+                .unwrap(),
+                runtime["subject_create_recovery_fences"]
+            );
+            drop(reopened);
+            let plan = FsRuntimeStore::migrate(&options).unwrap();
+            assert!(!plan.migration_required);
+            assert_eq!(plan.current_schema_version, 12);
+            assert_eq!(plan.target_schema_version, 12);
+            assert!(
+                !FsRuntimeStore::migration_plan(&options)
+                    .unwrap()
+                    .migration_required
+            );
+            assert_eq!(fs::read(&runtime_path).unwrap(), runtime_once);
+            assert_eq!(fs::read(&worker_path).unwrap(), worker_once);
+            assert_eq!(fs::read(&metadata_path).unwrap(), metadata);
+            assert_eq!(fs::read(&history_path).unwrap(), history);
+            assert!(
+                !migration_sibling(&root, "schema-v12-staging")
+                    .unwrap()
+                    .exists()
+            );
+            assert!(
+                !migration_sibling(&root, "pre-schema-v12-backup")
+                    .unwrap()
+                    .exists()
+            );
+        }
+    }
+
+    #[test]
+    fn schema_v11_migration_rejects_invalid_fences_without_replacing_them() {
+        for fences in [
+            serde_json::Value::Null,
+            serde_json::json!({"invalid-worker": {}}),
+        ] {
+            let temp = tempfile::tempdir().unwrap();
+            let root = temp.path().join("runtime");
+            write_empty_schema_v8_store(&root);
+            let path = root.join(RUNTIME_FILE);
+            let mut runtime: serde_json::Value = read_json(&path, "read fixture").unwrap();
+            runtime["schema_version"] = serde_json::json!(11);
+            runtime["subject_create_recovery_fences"] = fences;
+            atomic_write_json(&path, &runtime, "write invalid fence fixture").unwrap();
+            let before = fs::read(&path).unwrap();
+            assert!(FsRuntimeStore::open_or_create(root, "runtime-test").is_err());
+            assert_eq!(fs::read(&path).unwrap(), before);
+        }
+    }
+
     #[test]
     fn schema_v10_startup_migrates_stopped_sessions_atomically() {
         for corrupt in [false, true] {
@@ -2761,7 +3010,7 @@ mod tests {
                 assert_eq!(record["schema_version"], 10);
             } else {
                 drop(result.unwrap());
-                assert_eq!(stored_schema_version(&root), 11);
+                assert_eq!(stored_schema_version(&root), u64::from(SCHEMA_VERSION));
                 let manifest: serde_json::Value =
                     read_json(&session.join("session.json"), "test").unwrap();
                 assert_eq!(manifest["schema_version"], 4);
@@ -3151,7 +3400,7 @@ mod tests {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime-store");
         write_empty_schema_v8_store(&root);
-        let staging = migration_sibling(&root, "schema-v11-staging").unwrap();
+        let staging = migration_sibling(&root, "schema-v12-staging").unwrap();
         fs::create_dir_all(&staging).unwrap();
         fs::write(staging.join("partial"), b"partial").unwrap();
         let store = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap();
@@ -3166,7 +3415,7 @@ mod tests {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime-store");
         write_empty_schema_v8_store(&root);
-        let backup = migration_sibling(&root, "pre-schema-v11-backup").unwrap();
+        let backup = migration_sibling(&root, "pre-schema-v12-backup").unwrap();
         fs::rename(&root, &backup).unwrap();
         let store = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap();
 
@@ -3180,10 +3429,10 @@ mod tests {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime-store");
         write_empty_schema_v8_store(&root);
-        let staging = migration_sibling(&root, "schema-v11-staging").unwrap();
+        let staging = migration_sibling(&root, "schema-v12-staging").unwrap();
         copy_runtime_tree(&root, &root, &staging).unwrap();
         migrate_runtime_store_in_place(&staging, "test-runtime").unwrap();
-        let backup = migration_sibling(&root, "pre-schema-v11-backup").unwrap();
+        let backup = migration_sibling(&root, "pre-schema-v12-backup").unwrap();
         fs::rename(&root, &backup).unwrap();
         let store = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap();
 
@@ -3198,10 +3447,10 @@ mod tests {
         let temp = tempfile::tempdir().unwrap();
         let root = temp.path().join("runtime-store");
         write_empty_schema_v8_store(&root);
-        let staging = migration_sibling(&root, "schema-v11-staging").unwrap();
+        let staging = migration_sibling(&root, "schema-v12-staging").unwrap();
         fs::create_dir_all(&staging).unwrap();
         fs::write(staging.join(RUNTIME_FILE), b"{\"schema_version\":8}").unwrap();
-        let backup = migration_sibling(&root, "pre-schema-v11-backup").unwrap();
+        let backup = migration_sibling(&root, "pre-schema-v12-backup").unwrap();
         fs::rename(&root, &backup).unwrap();
 
         let store = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap();
@@ -3218,7 +3467,7 @@ mod tests {
         let root = temp.path().join("runtime-store");
         write_empty_schema_v8_store(&root);
         migrate_runtime_store_in_place(&root, "test-runtime").unwrap();
-        let backup = migration_sibling(&root, "pre-schema-v11-backup").unwrap();
+        let backup = migration_sibling(&root, "pre-schema-v12-backup").unwrap();
         write_empty_schema_v8_store(&backup);
         let store = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap();
 
@@ -3228,14 +3477,21 @@ mod tests {
 
     #[test]
     fn earlier_migration_backup_cannot_be_mistaken_for_a_fresh_runtime() {
-        let temp = tempfile::tempdir().unwrap();
-        let root = temp.path().join("runtime-store");
-        let backup = migration_sibling(&root, "pre-schema-v9-backup").unwrap();
-        write_empty_schema_v8_store(&backup);
-        let error = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap_err();
-        assert!(error.to_string().contains("explicit recovery"));
-        assert!(!root.exists());
-        assert!(backup.join(RUNTIME_FILE).exists());
+        for suffix in [
+            "pre-schema-v9-backup",
+            "pre-schema-v10-backup",
+            "schema-v11-staging",
+            "pre-schema-v11-backup",
+        ] {
+            let temp = tempfile::tempdir().unwrap();
+            let root = temp.path().join("runtime-store");
+            let backup = migration_sibling(&root, suffix).unwrap();
+            write_empty_schema_v8_store(&backup);
+            let error = FsRuntimeStore::open_or_create(root.clone(), "test-runtime").unwrap_err();
+            assert!(error.to_string().contains("explicit recovery"));
+            assert!(!root.exists());
+            assert!(backup.join(RUNTIME_FILE).exists());
+        }
     }
 
     #[test]

crates/worker-runtime/src/http_server.rs

@@ -2102,6 +2102,9 @@ fn required_runtime_permission(method: &Method, path: &str) -> Option<&'static s
     {
         return Some("workers:input");
     }
+    if path == "/v1/subject-create-recovery" && *method == Method::POST {
+        return Some("workers:delete");
+    }
     if path.ends_with("/stop")
         || path.ends_with("/cancel")
         || (path == "/v1/subject-release" && *method == Method::POST)
@@ -2973,6 +2976,19 @@ mod tests {
         }
     }
 
+    #[test]
+    fn subject_create_recovery_requires_worker_delete_permission() {
+        assert_eq!(
+            required_runtime_permission(&Method::POST, "/v1/subject-create-recovery"),
+            Some("workers:delete")
+        );
+        // Existing ordinary release authorization is unchanged.
+        assert_eq!(
+            required_runtime_permission(&Method::POST, "/v1/subject-release"),
+            Some("workers:stop")
+        );
+    }
+
     #[test]
     fn retention_routes_require_worker_delete_permission() {
         assert_eq!(
@@ -3197,6 +3213,14 @@ mod tests {
     struct AcceptingBackend;
 
     impl WorkerExecutionBackend for AcceptingBackend {
+        fn recover_subject_create(
+            &self,
+            _request: &runtime_api::SubjectCreateRecoveryRequest,
+        ) -> Result<(), String> {
+            // This test backend owns no real executions.
+            Ok(())
+        }
+
         fn backend_id(&self) -> &str {
             "http-test"
         }

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

@@ -406,6 +406,18 @@ impl runtime_api::RuntimeApi for RuntimeManagementApi {
         }
     }
 
+    async fn recover_subject_create(
+        &self,
+        request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, runtime_api::RuntimeApiError> {
+        let scope = material_scope(&self.state, &request.workspace_id)?;
+        self.state
+            .runtime
+            .recover_subject_create_scoped(&scope, request)
+            .map_err(RuntimeHttpRestError::runtime)
+            .map_err(api_error)
+    }
+
     async fn subject_release(
         &self,
         request: runtime_api::SubjectReleaseRequest,

crates/worker-runtime/src/http_server/tests/session_material_tests.rs

@@ -1086,6 +1086,54 @@ async fn canonical_http_rejects_replay_tampered_body_and_workspace_mismatch() {
     assert_eq!(unsigned.status(), StatusCode::UNAUTHORIZED);
 }
 
+#[tokio::test]
+async fn subject_create_recovery_signed_http_requires_delete_and_matching_workspace() {
+    let f = Fixture::new();
+    let identity = RuntimeIdentityMaterial::generate("recovery-key").unwrap();
+    let trust = workspace_trust(&identity);
+    let app = runtime_http_router_with_workspace_auth(
+        f.runtime.clone(),
+        None,
+        workspace_test_auth(vec![trust.clone()]),
+    );
+    let path = "/v1/subject-create-recovery";
+    let mut request = runtime_api::SubjectCreateRecoveryRequest {
+        workspace_id: f.scope.workspace_id.clone(),
+        worker_id: WorkerId::now_v7(),
+        create_fingerprint: "unpublished-create".into(),
+        operation_id: "recovery-1".into(),
+    };
+    for operation in ["workers:read", "workers:stop", "workers:create"] {
+        let body = serde_json::to_vec(&request).unwrap();
+        let mut claims = signed_claims(&trust, "POST", path, &body);
+        claims.operation = operation.into();
+        let response = signed_request(app.clone(), &identity, &claims, body, "workspace-a").await;
+        assert_eq!(response.status(), StatusCode::UNAUTHORIZED);
+    }
+    request.workspace_id = "other-workspace".into();
+    let body = serde_json::to_vec(&request).unwrap();
+    let claims = signed_claims(&trust, "POST", path, &body);
+    let response = signed_request(app.clone(), &identity, &claims, body, "workspace-a").await;
+    assert_eq!(response.status(), StatusCode::FORBIDDEN);
+    request.workspace_id = f.scope.workspace_id.clone();
+    for _ in 0..2 {
+        let body = serde_json::to_vec(&request).unwrap();
+        let claims = signed_claims(&trust, "POST", path, &body);
+        assert_eq!(claims.operation, "workers:delete");
+        let response = signed_request(app.clone(), &identity, &claims, body, "workspace-a").await;
+        assert_eq!(response.status(), StatusCode::OK);
+        let bytes = axum::body::to_bytes(response.into_body(), 16384)
+            .await
+            .unwrap();
+        let receipt: runtime_api::SubjectCreateRecoveryResponse =
+            serde_json::from_slice(&bytes).unwrap();
+        assert_eq!(receipt.worker_id, request.worker_id);
+        assert_eq!(receipt.workspace_id, request.workspace_id);
+        assert_eq!(receipt.create_fingerprint, request.create_fingerprint);
+        assert_eq!(receipt.operation_id, request.operation_id);
+    }
+}
+
 #[tokio::test]
 async fn explicit_material_erasure_requires_delete_capability_and_matching_scope() {
     let f = Fixture::new();

crates/worker-runtime/src/runtime.rs

@@ -355,7 +355,92 @@ impl Drop for WorkerOperationLease {
     }
 }
 
+/// Irreversible admission exclusion, independent of the removable Worker aggregate.
+/// A saved response is historical: retries still require a fresh backend absence check.
+#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub(crate) struct SubjectCreateRecoveryFence {
+    pub(crate) request: runtime_api::SubjectCreateRecoveryRequest,
+    pub(crate) response: Option<runtime_api::SubjectCreateRecoveryResponse>,
+}
+
 impl Runtime {
+    pub fn recover_subject_create_scoped(
+        &self,
+        scope: &RuntimeWorkspaceScope,
+        request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, RuntimeError> {
+        if request.workspace_id != scope.workspace_id
+            || request.create_fingerprint.trim().is_empty()
+            || request.operation_id.trim().is_empty()
+            || request.create_fingerprint.len() > 512
+            || request.operation_id.len() > 512
+        {
+            return Err(RuntimeError::InvalidRequest(
+                "invalid Subject create recovery identity".into(),
+            ));
+        }
+        let lease = self.worker_operation_lock(request.worker_id)?;
+        let _operation = lease.lock().map_err(|_| RuntimeError::StatePoisoned)?;
+        let backend = {
+            let mut state = self.lock()?;
+            state.ensure_running()?;
+            // A memory-only Runtime cannot issue durable proof.
+            if state.backend == RuntimeBackendKind::Memory {
+                return Err(RuntimeError::InvalidRequest(
+                    "Subject create recovery requires durable storage".into(),
+                ));
+            }
+            state.ensure_workspace_owner(scope, false)?;
+            if state.workers.contains_key(&request.worker_id) {
+                return Err(RuntimeError::InvalidRequest(
+                    "Subject create catalog record is still present".into(),
+                ));
+            }
+            if let Some(saved) = state.subject_create_recovery_fences.get(&request.worker_id) {
+                if saved.request != request {
+                    return Err(RuntimeError::InvalidRequest(
+                        "different Subject create recovery is retained".into(),
+                    ));
+                }
+            } else {
+                state.ensure_workspace_owner(scope, true)?;
+                state.subject_create_recovery_fences.insert(
+                    request.worker_id,
+                    SubjectCreateRecoveryFence {
+                        request: request.clone(),
+                        response: None,
+                    },
+                );
+            }
+            // Never roll back admission exclusion, including uncertain persistence failures.
+            // Persist it before even checking backend availability.
+            state.persist_runtime_snapshot()?;
+            state.execution_backend.clone().ok_or_else(|| {
+                RuntimeError::InvalidRequest(
+                    "Subject create execution absence cannot be confirmed".into(),
+                )
+            })?
+        };
+        backend
+            .recover_subject_create(&request)
+            .map_err(RuntimeError::InvalidRequest)?;
+        let response = runtime_api::SubjectCreateRecoveryResponse {
+            workspace_id: request.workspace_id.clone(),
+            worker_id: request.worker_id,
+            create_fingerprint: request.create_fingerprint.clone(),
+            operation_id: request.operation_id.clone(),
+        };
+        let mut state = self.lock()?;
+        state
+            .subject_create_recovery_fences
+            .get_mut(&request.worker_id)
+            .expect("recovery fence retained under Worker operation lock")
+            .response = Some(response.clone());
+        state.persist_runtime_snapshot()?;
+        Ok(response)
+    }
+
     /// Reserves the actual Controller while the Host reconciles final saved Journal material.
     /// A Runtime restart loses proof of a reservation, never reconstructs it from Worker status.
     pub fn subject_release_scoped(
@@ -1262,6 +1347,8 @@ impl Runtime {
         let _operation_guard = operation_lock
             .lock()
             .map_err(|_| RuntimeError::StatePoisoned)?;
+        self.lock()?
+            .ensure_subject_create_not_fenced(request.worker_id)?;
         if let Some(existing) = self.existing_worker_for_create(&request, scope)? {
             return Ok(existing);
         }
@@ -5482,6 +5569,7 @@ struct RuntimeState {
     #[cfg(feature = "fs-store")]
     next_diagnostic_id: u64,
     workers: BTreeMap<WorkerId, WorkerRecord>,
+    subject_create_recovery_fences: BTreeMap<WorkerId, SubjectCreateRecoveryFence>,
     subject_release_operations: BTreeMap<
         (WorkerId, String),
         (
@@ -5519,6 +5607,7 @@ impl RuntimeState {
             #[cfg(feature = "fs-store")]
             next_diagnostic_id: 1,
             workers: BTreeMap::new(),
+            subject_create_recovery_fences: BTreeMap::new(),
             subject_release_operations: BTreeMap::new(),
             workdir_operations: BTreeMap::new(),
             workspace_owners: BTreeMap::new(),
@@ -5551,6 +5640,7 @@ impl RuntimeState {
             #[cfg(feature = "fs-store")]
             next_diagnostic_id: 1,
             workers: BTreeMap::new(),
+            subject_create_recovery_fences: BTreeMap::new(),
             subject_release_operations: BTreeMap::new(),
             workdir_operations: BTreeMap::new(),
             workspace_owners: BTreeMap::new(),
@@ -5697,6 +5787,7 @@ impl RuntimeState {
             workspace_backend_resource_clients: BTreeMap::new(),
             next_diagnostic_id,
             workers,
+            subject_create_recovery_fences: persisted.subject_create_recovery_fences,
             subject_release_operations: persisted
                 .subject_releases
                 .into_iter()
@@ -5727,6 +5818,7 @@ impl RuntimeState {
     #[cfg(feature = "fs-store")]
     fn persisted_state(&self) -> PersistedRuntimeState {
         PersistedRuntimeState {
+            subject_create_recovery_fences: self.subject_create_recovery_fences.clone(),
             subject_releases: self.subject_release_operations.values().cloned().collect(),
             display_name: self.display_name.clone(),
             status: self.status,
@@ -5871,11 +5963,21 @@ impl RuntimeState {
         Ok(())
     }
 
+    fn ensure_subject_create_not_fenced(&self, worker_id: WorkerId) -> Result<(), RuntimeError> {
+        if self.subject_create_recovery_fences.contains_key(&worker_id) {
+            return Err(RuntimeError::InvalidRequest(
+                "Subject create identity is irreversibly fenced".into(),
+            ));
+        }
+        Ok(())
+    }
+
     fn ensure_subject_not_reserved(
         &self,
         worker: &WorkerRef,
         allow_disconnected: bool,
     ) -> Result<(), RuntimeError> {
+        self.ensure_subject_create_not_fenced(worker.worker_id)?;
         use runtime_api::SubjectReleaseState as State;
         if self
             .subject_release_operations
@@ -5945,6 +6047,13 @@ impl RuntimeState {
     }
 
     fn forget_workspace_owner_if_unused(&mut self, workspace_id: &str) -> bool {
+        if self
+            .subject_create_recovery_fences
+            .values()
+            .any(|fence| fence.request.workspace_id == workspace_id)
+        {
+            return false;
+        }
         if self
             .workers
             .values()
@@ -8293,6 +8402,8 @@ mod tests {
 
     #[derive(Default)]
     struct TestExecutionBackend {
+        recovery_absent: AtomicBool,
+        recovery_checks: std::sync::atomic::AtomicUsize,
         spawn_result: Mutex<Option<WorkerExecutionSpawnResult>>,
         job_bindings: Mutex<Vec<Option<crate::catalog::BackendJobExecutionBinding>>>,
         dispatch_result: Mutex<Option<WorkerExecutionResult>>,
@@ -8367,6 +8478,18 @@ mod tests {
     }
 
     impl WorkerExecutionBackend for TestExecutionBackend {
+        fn recover_subject_create(
+            &self,
+            _request: &runtime_api::SubjectCreateRecoveryRequest,
+        ) -> Result<(), String> {
+            self.recovery_checks.fetch_add(1, Ordering::SeqCst);
+            if self.recovery_absent.load(Ordering::SeqCst) {
+                Ok(())
+            } else {
+                Err("execution absence is unconfirmed".into())
+            }
+        }
+
         fn subject_release(
             &self,
             request: &runtime_api::SubjectReleaseRequest,
@@ -13098,7 +13221,7 @@ mod tests {
         })
         .unwrap_err();
         assert!(error.to_string().contains(
-            "unsupported Runtime store schema version 2; expected 6, 7, 8, 9, 10, or 11"
+            "unsupported Runtime store schema version 2; expected 6, 7, 8, 9, 10, 11, or 12"
         ));
 
         let _ = std::fs::remove_dir_all(root);
@@ -13140,7 +13263,7 @@ mod tests {
         let worker_aggregate: serde_json::Value =
             serde_json::from_slice(&std::fs::read(worker_store_dir.join("worker.json")).unwrap())
                 .unwrap();
-        assert_eq!(worker_aggregate["schema_version"], serde_json::json!(11));
+        assert_eq!(worker_aggregate["schema_version"], serde_json::json!(12));
         assert_eq!(worker_aggregate["status"], serde_json::json!("stopped"));
         assert_eq!(
             worker_aggregate["execution_state"]["state"],
@@ -13280,6 +13403,213 @@ mod tests {
         let _ = std::fs::remove_dir_all(root);
     }
 
+    #[cfg(feature = "fs-store")]
+    #[test]
+    fn subject_create_recovery_is_durable_exact_and_rechecks_absence() {
+        let root = tempfile::tempdir().unwrap();
+        let options = crate::fs_store::FsRuntimeStoreOptions::new(root.path())
+            .with_runtime_id("test-runtime");
+        let scoped = scope("workspace-a", "server-a");
+        let backend = Arc::new(TestExecutionBackend::default());
+        backend.recovery_absent.store(true, Ordering::SeqCst);
+        let runtime =
+            Runtime::with_fs_store_and_execution_backend(options.clone(), backend.clone()).unwrap();
+        let create = scoped_task_request("unpublished create", "workspace-a");
+        let request = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: scoped.workspace_id.clone(),
+            worker_id: create.worker_id,
+            create_fingerprint: create.create_fingerprint.clone(),
+            operation_id: "recovery-1".into(),
+        };
+        runtime.store_config_bundle(test_bundle()).unwrap();
+        *backend.spawn_result.lock().unwrap() = Some(WorkerExecutionSpawnResult::RolledBack(
+            WorkerExecutionResult::rejected(
+                WorkerExecutionOperation::Spawn,
+                "create failed before publication",
+            ),
+        ));
+        assert!(
+            runtime
+                .create_worker_scoped(&scoped, create.clone())
+                .is_err()
+        );
+        assert!(
+            !runtime
+                .lock()
+                .unwrap()
+                .workers
+                .contains_key(&create.worker_id)
+        );
+        let response = runtime
+            .recover_subject_create_scoped(&scoped, request.clone())
+            .unwrap();
+        assert_eq!(response.worker_id, request.worker_id);
+        assert!(
+            runtime
+                .create_worker_scoped(&scoped, create.clone())
+                .is_err()
+        );
+        assert!(
+            !runtime
+                .lock()
+                .unwrap()
+                .forget_workspace_owner_if_unused("workspace-a")
+        );
+        assert_eq!(
+            runtime
+                .lock()
+                .unwrap()
+                .workspace_owners
+                .get("workspace-a")
+                .map(String::as_str),
+            Some("server-a")
+        );
+        assert_eq!(
+            runtime
+                .recover_subject_create_scoped(&scoped, request.clone())
+                .unwrap(),
+            response
+        );
+        assert_eq!(backend.recovery_checks.load(Ordering::SeqCst), 2);
+        for mut mismatch in [request.clone(), request.clone(), request.clone()]
+            .into_iter()
+            .enumerate()
+        {
+            match mismatch.0 {
+                0 => mismatch.1.create_fingerprint.push_str("-wrong"),
+                1 => mismatch.1.operation_id.push_str("-wrong"),
+                _ => mismatch.1.workspace_id.push_str("-wrong"),
+            }
+            assert!(
+                runtime
+                    .recover_subject_create_scoped(&scoped, mismatch.1)
+                    .is_err()
+            );
+        }
+        assert!(
+            runtime
+                .recover_subject_create_scoped(
+                    &scope("workspace-a", "other-server"),
+                    request.clone()
+                )
+                .is_err()
+        );
+        drop(runtime);
+        let backend = Arc::new(TestExecutionBackend::default());
+        let reopened =
+            Runtime::with_fs_store_and_execution_backend(options, backend.clone()).unwrap();
+        assert_eq!(*backend.restore_count.lock().unwrap(), 0);
+        assert!(reopened.create_worker_scoped(&scoped, create).is_err());
+        assert!(
+            reopened
+                .recover_subject_create_scoped(&scoped, request.clone())
+                .is_err(),
+            "saved success cannot replace fresh absence proof"
+        );
+        backend.recovery_absent.store(true, Ordering::SeqCst);
+        assert_eq!(
+            reopened
+                .recover_subject_create_scoped(&scoped, request)
+                .unwrap(),
+            response
+        );
+        assert_eq!(backend.recovery_checks.load(Ordering::SeqCst), 2);
+    }
+
+    #[cfg(feature = "fs-store")]
+    #[test]
+    fn subject_create_recovery_unknown_backend_retains_fence_and_owner() {
+        let root = tempfile::tempdir().unwrap();
+        let options = crate::fs_store::FsRuntimeStoreOptions::new(root.path())
+            .with_runtime_id("test-runtime");
+        let scoped = scope("workspace-a", "server-a");
+        let runtime = Runtime::with_fs_store(options.clone()).unwrap();
+        let create = scoped_task_request("unpublished create", "workspace-a");
+        let request = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: scoped.workspace_id.clone(),
+            worker_id: create.worker_id,
+            create_fingerprint: create.create_fingerprint.clone(),
+            operation_id: "recovery-unknown".into(),
+        };
+        assert!(
+            runtime
+                .recover_subject_create_scoped(&scoped, request.clone())
+                .is_err()
+        );
+        assert!(
+            runtime
+                .lock()
+                .unwrap()
+                .subject_create_recovery_fences
+                .get(&request.worker_id)
+                .unwrap()
+                .response
+                .is_none()
+        );
+        drop(runtime);
+        let backend = Arc::new(TestExecutionBackend::default());
+        let reopened =
+            Runtime::with_fs_store_and_execution_backend(options, backend.clone()).unwrap();
+        assert!(reopened.create_worker_scoped(&scoped, create).is_err());
+        assert!(
+            reopened
+                .recover_subject_create_scoped(
+                    &scope("workspace-a", "other-server"),
+                    request.clone()
+                )
+                .is_err()
+        );
+        assert!(
+            reopened
+                .recover_subject_create_scoped(&scoped, request.clone())
+                .is_err()
+        );
+        backend.recovery_absent.store(true, Ordering::SeqCst);
+        reopened
+            .recover_subject_create_scoped(&scoped, request)
+            .unwrap();
+    }
+
+    #[cfg(feature = "fs-store")]
+    #[test]
+    fn subject_create_recovery_rejects_existing_catalog_even_when_stopped() {
+        let root = tempfile::tempdir().unwrap();
+        let backend = Arc::new(TestExecutionBackend::default());
+        backend.recovery_absent.store(true, Ordering::SeqCst);
+        let runtime = Runtime::with_fs_store_and_execution_backend(
+            crate::fs_store::FsRuntimeStoreOptions::new(root.path())
+                .with_runtime_id("test-runtime"),
+            backend.clone(),
+        )
+        .unwrap();
+        runtime.store_config_bundle(test_bundle()).unwrap();
+        let scoped = scope("workspace-a", "server-a");
+        let create = scoped_task_request("existing subject", "workspace-a");
+        let worker = runtime
+            .create_worker_scoped(&scoped, create.clone())
+            .unwrap();
+        runtime.stop_worker(&worker.worker_ref, None).unwrap();
+        let request = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: scoped.workspace_id.clone(),
+            worker_id: create.worker_id,
+            create_fingerprint: create.create_fingerprint.clone(),
+            operation_id: "not-unpublished".into(),
+        };
+        assert!(
+            runtime
+                .recover_subject_create_scoped(&scoped, request)
+                .is_err()
+        );
+        assert_eq!(backend.recovery_checks.load(Ordering::SeqCst), 0);
+        assert!(
+            runtime
+                .lock()
+                .unwrap()
+                .subject_create_recovery_fences
+                .is_empty()
+        );
+    }
+
     // Persistence component test; real Controller reservation has separate coverage.
     #[cfg(feature = "fs-store")]
     #[test]
@@ -14164,7 +14494,7 @@ mod tests {
         );
         let migrated_aggregate: serde_json::Value =
             serde_json::from_slice(&std::fs::read(&worker_path).unwrap()).unwrap();
-        assert_eq!(migrated_aggregate["schema_version"], serde_json::json!(11));
+        assert_eq!(migrated_aggregate["schema_version"], serde_json::json!(12));
         assert!(migrated_aggregate.get("request").is_none());
         assert!(migrated_aggregate.get("execution").is_none());
         assert_eq!(

crates/worker-runtime/src/worker_backend.rs

@@ -3369,6 +3369,42 @@ where
         &self.backend_id
     }
 
+    fn recover_subject_create(
+        &self,
+        request: &runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<(), String> {
+        let worker_ref = WorkerRef::new(request.worker_id);
+        let lock = self.worker_lock(&worker_ref)?;
+        let _operation = lock.lock().map_err(|_| "worker operation lock poisoned")?;
+        if self
+            .runtime
+            .lock()
+            .map_err(|_| "adapter runtime poisoned")?
+            .is_none()
+        {
+            return Err("execution backend is unavailable".into());
+        }
+        if self
+            .pending_restores
+            .lock()
+            .map_err(|_| "restore registry poisoned")?
+            .contains_key(&worker_ref)
+        {
+            return Err("Subject create has a pending execution candidate".into());
+        }
+        if self
+            .workers
+            .lock()
+            .map_err(|_| "worker registry poisoned")?
+            .contains_key(&worker_ref)
+        {
+            // Even a closed protocol or stopped snapshot is residual execution,
+            // not absence. Recovery does not stop or remove anything.
+            return Err("Subject create execution is still present".into());
+        }
+        Ok(())
+    }
+
     fn subject_release(
         &self,
         request: &runtime_api::SubjectReleaseRequest,
@@ -6069,6 +6105,48 @@ mod tests {
         assert!(backend.workers.lock().unwrap().is_empty());
     }
 
+    #[test]
+    fn subject_create_recovery_checks_actual_absence_and_pending_candidates() {
+        let root = tempfile::tempdir().unwrap();
+        let backend = WorkerRuntimeExecutionBackend::new(
+            ProfileRuntimeWorkerFactory::new(root.path())
+                .with_runtime_store_dir(root.path().join("runtime")),
+        )
+        .unwrap();
+        let worker_ref = WorkerRef::new(WorkerId::now_v7());
+        let recovery = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: "ws-test".into(),
+            worker_id: worker_ref.worker_id,
+            create_fingerprint: "failed-create".into(),
+            operation_id: "recovery-1".into(),
+        };
+        backend.recover_subject_create(&recovery).unwrap();
+        // No Controller exists, but unresolved task ownership is not absence.
+        backend.pending_restores.lock().unwrap().insert(
+            worker_ref.clone(),
+            Arc::new(PendingRuntimeRestore {
+                request: direct_restore_request(worker_ref.clone()),
+                task: tokio::sync::Mutex::new(None),
+                result: Mutex::new(None),
+                failure: Mutex::new(Some("task completion unknown".into())),
+            }),
+        );
+        assert!(
+            backend
+                .recover_subject_create(&recovery)
+                .unwrap_err()
+                .contains("pending")
+        );
+        assert!(
+            backend
+                .pending_restores
+                .lock()
+                .unwrap()
+                .contains_key(&worker_ref)
+        );
+        assert!(backend.workers.lock().unwrap().is_empty());
+    }
+
     #[test]
     fn pending_restore_join_failure_never_becomes_absence_on_stop_retry() {
         let root = tempfile::tempdir().unwrap();
@@ -6127,6 +6205,18 @@ mod tests {
     fn stop_joins_healthy_self_terminated_controller_with_closed_endpoint() {
         let root = tempfile::tempdir().unwrap();
         let (backend, request) = profile_stop_fixture(root.path());
+        let recovery = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: "ws-test".into(),
+            worker_id: request.worker_ref.worker_id,
+            create_fingerprint: "failed-create".into(),
+            operation_id: "recovery-1".into(),
+        };
+        assert!(
+            backend
+                .recover_subject_create(&recovery)
+                .unwrap_err()
+                .contains("still present")
+        );
         let execution = backend
             .workers
             .lock()
@@ -6167,8 +6257,16 @@ mod tests {
             })
             .unwrap();
         assert!(!execution.shutdown_requested.load(Ordering::Acquire));
+        assert!(
+            backend
+                .recover_subject_create(&recovery)
+                .unwrap_err()
+                .contains("still present"),
+            "closed Controller is still residual execution"
+        );
         assert!(backend.stop_worker(&request.worker_ref).is_accepted());
         assert!(backend.workers.lock().unwrap().is_empty());
+        backend.recover_subject_create(&recovery).unwrap();
         assert!(execution.tasks.tasks.lock().unwrap().is_empty());
         assert!(backend.stop_worker(&request.worker_ref).is_accepted());
     }

crates/worker/src/worker.rs

@@ -11298,6 +11298,7 @@ mod build_summary_prompt_tests {
     enum AttributionOutcome {
         Success,
         Rejected,
+        Unauthorized,
         TransportUnknown,
     }
 
@@ -11344,6 +11345,16 @@ mod build_summary_prompt_tests {
                 AttributionOutcome::TransportUnknown => Err(WorkspaceClientError::Request(
                     "connection closed after request send".to_string(),
                 )),
+                AttributionOutcome::Unauthorized => Ok(WorkspaceResponse {
+                    status: 401,
+                    body: serde_json::to_string(&server_api::RepositoryApiError::new(
+                        401,
+                        "Unauthorized",
+                        "invalid runtime request proof",
+                        Vec::new(),
+                    ))
+                    .unwrap(),
+                }),
                 AttributionOutcome::Rejected => Ok(WorkspaceResponse {
                     status: 503,
                     body: "temporarily unavailable".to_string(),
@@ -11576,6 +11587,7 @@ mod build_summary_prompt_tests {
         for (first, expected_state) in [
             (AttributionOutcome::TransportUnknown, "outcome_unknown"),
             (AttributionOutcome::Rejected, "failed"),
+            (AttributionOutcome::Unauthorized, "failed"),
         ] {
             let client = Arc::new(AttributionLifecycleClient::new([
                 first,
@@ -11589,6 +11601,15 @@ mod build_summary_prompt_tests {
                 )
                 .unwrap_err();
             assert!(error.to_string().contains(expected_state));
+            if matches!(first, AttributionOutcome::Unauthorized) {
+                assert!(error.to_string().contains("rejected with HTTP 401"));
+                assert!(!error.to_string().contains("outcome_unknown"));
+                assert_eq!(
+                    client.requests.lock().unwrap().len(),
+                    1,
+                    "no automatic retry after auth rejection"
+                );
+            }
             let incomplete = store
                 .read_by_name(&worker.manifest().worker.name)
                 .unwrap()

crates/workspace-server/src/assignment_tracking_tests.rs

@@ -475,7 +475,8 @@ fn tracking_store_real_schema91_to92_retains_queue_cases_principals_audit_receip
                     (98, "Assignment-independent implementation reports".into()),
                     (99, "Explicit hosting administrators and self-registration policy".into()),
                     (100, "Provisional users and scoped passkey activation".into()),
-                    (101, "Durable explicit Subject body release".into())
+                    (101, "Durable explicit Subject body release".into()),
+                    (102, "Durable failed Subject creation recovery".into())
                 ]
             );
             for (table, rows) in retained.iter().zip(original.iter()) {

crates/workspace-server/src/hosts.rs

@@ -1288,6 +1288,13 @@ pub trait WorkspaceWorkerRuntime: Send + Sync {
         })
     }
 
+    fn recover_subject_create(
+        &self,
+        _request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, String> {
+        Err("Subject create recovery fencing is unavailable on this Runtime".into())
+    }
+
     fn subject_release(
         &self,
         _request: runtime_api::SubjectReleaseRequest,
@@ -2251,6 +2258,21 @@ impl RuntimeRegistry {
         runtime.worker_session_attachment(workspace_id, &worker_ref, session_id, attachment_id)
     }
 
+    pub fn recover_subject_create(
+        &self,
+        runtime_id: &str,
+        request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, RuntimeRegistryError> {
+        validate_backend_identifier("runtime_id", runtime_id)?;
+        self.runtime(runtime_id)?
+            .recover_subject_create(request)
+            .map_err(|message| RuntimeRegistryError::RuntimeOperationFailed {
+                runtime_id: runtime_id.into(),
+                code: "recover_subject_create_unproven".into(),
+                message,
+            })
+    }
+
     pub fn subject_release(
         &self,
         runtime_id: &str,
@@ -3319,6 +3341,19 @@ impl WorkspaceWorkerRuntime for EmbeddedWorkerRuntime {
             })
     }
 
+    fn recover_subject_create(
+        &self,
+        request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, String> {
+        if request.workspace_id != self.workspace_id {
+            return Err("Subject create recovery Workspace mismatch".into());
+        }
+        let scope = RuntimeWorkspaceScope::new(&request.workspace_id, "embedded-backend");
+        self.runtime
+            .recover_subject_create_scoped(&scope, request)
+            .map_err(|e| e.to_string())
+    }
+
     fn subject_release(
         &self,
         request: runtime_api::SubjectReleaseRequest,
@@ -4164,7 +4199,10 @@ fn workspace_runtime_operation(method: &str, path_and_query: &str) -> &'static s
     if path.ends_with("/completions") {
         return "workers:read";
     }
-    if path.contains("/retention/") || (path.starts_with("/v1/workers/") && method == "DELETE") {
+    if path.contains("/retention/")
+        || (path.starts_with("/v1/workers/") && method == "DELETE")
+        || (path == "/v1/subject-create-recovery" && method == "POST")
+    {
         return "workers:delete";
     }
     if path.starts_with("/v1/session-public/") && method == "POST" {
@@ -5561,6 +5599,18 @@ impl WorkspaceWorkerRuntime for RemoteWorkerRuntime {
         })
     }
 
+    fn recover_subject_create(
+        &self,
+        request: runtime_api::SubjectCreateRecoveryRequest,
+    ) -> Result<runtime_api::SubjectCreateRecoveryResponse, String> {
+        self.run_runtime_api(
+            self.request_timeout,
+            MAX_REMOTE_RUNTIME_RESPONSE_BYTES,
+            move |client| async move { client.recover_subject_create(request).await },
+        )
+        .map_err(|diagnostic| diagnostic.message)
+    }
+
     fn subject_release(
         &self,
         request: runtime_api::SubjectReleaseRequest,
@@ -8377,6 +8427,51 @@ mod tests {
         );
     }
 
+    #[test]
+    fn remote_subject_create_recovery_uses_generated_management_transport() {
+        assert_eq!(
+            workspace_runtime_operation("POST", "/v1/subject-create-recovery"),
+            "workers:delete"
+        );
+        let request = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: "workspace-test".into(),
+            worker_id: EmbeddedWorkerId::now_v7(),
+            create_fingerprint: "sha256:failed-create".into(),
+            operation_id: "recover-1".into(),
+        };
+        // Remote adapter transport only; signed Runtime authorization is tested
+        // in worker-runtime's actual HTTP router suite.
+        let (base_url, server) = serve_mock_http(vec![mock_response(
+            "POST",
+            "/v1/subject-create-recovery",
+            true,
+            200,
+            serde_json::to_string(&request).unwrap(),
+        )]);
+        let registry = RuntimeRegistry::new(Vec::new());
+        registry.register(
+            RemoteWorkerRuntime::new(
+                RemoteRuntimeConfig::new(
+                    "remote:recover",
+                    "Recovery transport fixture",
+                    base_url,
+                    Some("secret-token".into()),
+                ),
+                "workspace-test".into(),
+                "http://127.0.0.1:8787".into(),
+            )
+            .unwrap(),
+        );
+        let response = registry
+            .recover_subject_create("remote:recover", request.clone())
+            .unwrap();
+        assert_eq!(response.workspace_id, request.workspace_id);
+        assert_eq!(response.worker_id, request.worker_id);
+        assert_eq!(response.create_fingerprint, request.create_fingerprint);
+        assert_eq!(response.operation_id, request.operation_id);
+        server.join().unwrap();
+    }
+
     #[test]
     fn remote_session_erasure_uses_the_generated_management_transport() {
         // The loopback fixture substitutes only the remote Runtime. This covers

crates/workspace-server/src/implementation_result_tests.rs

@@ -381,6 +381,8 @@ async fn implementation_result_v97_to_v98_preserves_legacy_bytes_receipts_and_no
             DROP TABLE ticket_implementation_results;
             DROP TABLE ticket_implementation_delegations;
             DROP TABLE merge_request_source_observations;
+            DROP TABLE subject_create_recoveries;
+            UPDATE __yoi_schema_migrations SET version=101 WHERE version=102;
             DROP TABLE subject_worker_releases;
             UPDATE __yoi_schema_migrations SET version=100 WHERE version=101;
             DROP TABLE user_activations;
@@ -515,6 +517,8 @@ async fn mr_notification_upgrade_recovers_events_not_comments_and_retains_unknow
         // Reconstruct the prior composition before its new business capture existed.
         conn.execute_batch("DROP TRIGGER notification_mr_event; DROP TRIGGER notification_mr_source; DROP TRIGGER notification_implementation_result; DROP TRIGGER IF EXISTS implementation_delegation_capture;
             DROP TABLE ticket_implementation_results; DROP TABLE ticket_implementation_delegations; DROP TABLE merge_request_source_observations;
+            DROP TABLE subject_create_recoveries;
+            UPDATE __yoi_schema_migrations SET version=101 WHERE version=102;
             DROP TABLE subject_worker_releases;
             UPDATE __yoi_schema_migrations SET version=100 WHERE version=101;
             DROP TABLE user_activations;
@@ -675,7 +679,9 @@ async fn mr_event_actors_preserve_self_suppression_in_capture_and_backfill() {
             if historical {
                 conn.execute_batch("DROP TRIGGER notification_mr_event; DROP TRIGGER notification_mr_source; DROP TRIGGER notification_implementation_result; DROP TRIGGER IF EXISTS implementation_delegation_capture;
                     DROP TABLE ticket_implementation_results; DROP TABLE ticket_implementation_delegations; DROP TABLE merge_request_source_observations;
-                    DROP TABLE subject_worker_releases;
+                    DROP TABLE subject_create_recoveries;
+            UPDATE __yoi_schema_migrations SET version=101 WHERE version=102;
+            DROP TABLE subject_worker_releases;
             UPDATE __yoi_schema_migrations SET version=100 WHERE version=101;
             DROP TABLE user_activations;
             UPDATE __yoi_schema_migrations SET version=99 WHERE version=100;

crates/workspace-server/src/latest_schema.sql

@@ -699,6 +699,19 @@ CREATE TABLE worker_create_reservations (
             FOREIGN KEY (workspace_id) REFERENCES workspaces(workspace_id) ON DELETE CASCADE,
             CHECK ((singleton_key IS NULL) = (singleton_generation IS NULL))
         );
+CREATE TABLE subject_create_recoveries (
+    workspace_id TEXT NOT NULL,
+    runtime_id TEXT NOT NULL,
+    worker_id TEXT NOT NULL,
+    singleton_key TEXT NOT NULL,
+    singleton_generation INTEGER NOT NULL CHECK (singleton_generation > 0),
+    operation_id TEXT NOT NULL CHECK (length(operation_id) > 0),
+    create_fingerprint TEXT NOT NULL,
+    completed INTEGER NOT NULL DEFAULT 0 CHECK (completed IN (0,1)),
+    PRIMARY KEY (workspace_id, operation_id),
+    UNIQUE (workspace_id, runtime_id, worker_id, singleton_key, singleton_generation),
+    FOREIGN KEY (workspace_id, worker_id) REFERENCES worker_create_reservations(workspace_id, worker_id) ON DELETE CASCADE
+);
 CREATE TABLE subject_worker_releases (
     workspace_id TEXT NOT NULL,
     runtime_id TEXT NOT NULL,

crates/workspace-server/src/server.rs

@@ -2934,10 +2934,10 @@ fn repository_api_rejection(path: &str, status: StatusCode, message: &str) -> Re
     if path.starts_with("/api/w/") && path.contains("/drive/") {
         return (status, Json(drive::failure(status.as_u16()))).into_response();
     }
-    if path == "/api/repositories"
-        || path.starts_with("/api/repositories/")
-        || (path.starts_with("/api/w/") && path.contains("/repositories"))
-    {
+    // Generated Workspace API clients share this error envelope, not only
+    // repository endpoints. A bare 401 here becomes a response-decode failure
+    // and incorrectly turns a known auth rejection into an unknown outcome.
+    if path.starts_with("/api/") {
         return (
             status,
             Json(server_api::RepositoryApiError::new(
@@ -5484,6 +5484,12 @@ impl WorkspaceApi {
         let subject = singleton_key
             .as_deref()
             .and_then(|key| key.strip_prefix(SUBJEKTIV_SINGLETON_PREFIX));
+        if let Some(subject) = subject
+            && let Err(error) = subjektiv_journal::recover_failed_subject_create(self, subject)
+        {
+            let diagnostics = cleanup_spawn_created_workdirs(self, &spawned_workdir_ids);
+            return Err(ApiError::with_diagnostics(error, diagnostics));
+        }
         let subject_gate = subject
             .map(|s| subjektiv_journal::lifecycle_lock(self, s))
             .transpose()?;
@@ -22372,15 +22378,12 @@ async fn scoped_record_subjektiv_session(
         .run(move || {
             validate_workspace_scope(&api, &path.workspace_id)?;
             let attribution = if request.create_if_missing {
-                let attribution = subjektiv_session_attribution(
+                subjektiv_session_attribution(
                     &api,
                     &path.workspace_id,
                     &context,
                     &request.session_id,
-                )?;
-                open_subjektiv_store(&api)?
-                    .record_session_attribution(attribution)
-                    .map_err(|error| Error::InvalidInput(error.to_string()))?
+                )?
             } else {
                 require_existing_subjektiv_session_attribution(
                     &api,
@@ -47939,6 +47942,9 @@ mod tests {
         >,
         materializer: worker_runtime::working_directory::RuntimeGitMaterializer,
         spawn_failure: std::sync::Mutex<Option<String>>,
+        recovery_failure: std::sync::Mutex<Option<String>>,
+        recovery_requests: std::sync::Mutex<Vec<runtime_api::SubjectCreateRecoveryRequest>>,
+        recovery_hook: std::sync::Mutex<Option<Box<dyn FnOnce() + Send>>>,
         input_failure: std::sync::Mutex<Option<String>>,
         cancel_failure: std::sync::Mutex<Option<String>>,
         input_hook: std::sync::Mutex<
@@ -47974,6 +47980,9 @@ mod tests {
                     std::env::temp_dir().join(unique),
                 ),
                 spawn_failure: std::sync::Mutex::new(None),
+                recovery_failure: std::sync::Mutex::new(None),
+                recovery_requests: std::sync::Mutex::new(Vec::new()),
+                recovery_hook: std::sync::Mutex::new(None),
                 input_failure: std::sync::Mutex::new(None),
                 cancel_failure: std::sync::Mutex::new(None),
                 input_hook: std::sync::Mutex::new(None),
@@ -48068,6 +48077,29 @@ mod tests {
             "deterministic-workspace-server-test"
         }
 
+        fn recover_subject_create(
+            &self,
+            request: &runtime_api::SubjectCreateRecoveryRequest,
+        ) -> std::result::Result<(), String> {
+            self.recovery_requests.lock().unwrap().push(request.clone());
+            if let Some(hook) = self.recovery_hook.lock().unwrap().take() {
+                hook();
+            }
+            if let Some(error) = self.recovery_failure.lock().unwrap().take() {
+                return Err(error);
+            }
+            if self
+                .contexts
+                .lock()
+                .unwrap()
+                .keys()
+                .any(|worker| worker.worker_id == request.worker_id)
+            {
+                return Err("actual execution remains".into());
+            }
+            Ok(())
+        }
+
         fn create_working_directory(
             &self,
             request: &worker_runtime::catalog::WorkingDirectoryRequest,

crates/workspace-server/src/server/subjektiv_journal.rs

@@ -346,6 +346,11 @@ pub(super) fn bind_host_session(
             }
         }
     }
+    // Publication (including attribution) shares the recovery gate. A callback
+    // authenticated before a recovery intent must not leave a late attribution.
+    let attr = store
+        .record_session_attribution(attr)
+        .map_err(|error| Error::InvalidInput(error.to_string()))?;
     let writer = store
         .claim_journal_writer(&attr.subject_id, &attr.runtime_id, &attr.worker_id)
         .map_err(domain)?;
@@ -386,6 +391,93 @@ pub(super) fn check_start_allowed(api: &WorkspaceApi, subject: &str) -> Result<(
     Ok(())
 }
 
+/// Explicit creation can reconcile an earlier failed, never-published body.
+/// No missing catalog/status observation is treated as a Runtime fencing receipt.
+pub(super) fn recover_failed_subject_create(api: &WorkspaceApi, subject: &str) -> Result<()> {
+    let continuation = lifecycle_lock(api, &format!("release-continuation:{subject}"))?;
+    let _continuation = continuation
+        .lock()
+        .map_err(|_| Error::Store("release continuation poisoned".into()))?;
+    let gate = lifecycle_lock(api, subject)?;
+    let subject_key = crate::subjektiv::subject_worker_singleton_key(subject)
+        .map_err(|e| Error::InvalidInput(e.to_string()))?;
+    let store = store(api)?;
+    let (lease, request) = {
+        let _guard = gate
+            .lock()
+            .map_err(|_| Error::Store("Journal lifecycle lock poisoned".into()))?;
+        let Some(lease) = api
+            .config_store
+            .current_worker_singleton_owner(api.workspace_id(), &subject_key)?
+        else {
+            return Ok(());
+        };
+        if api
+            .config_store
+            .subject_create_recovery_candidate(api.workspace_id(), &lease)?
+            .is_none()
+        {
+            return Ok(());
+        }
+        if !store
+            .journal_body_is_unpublished(subject, &lease.worker.runtime_id, &lease.worker.worker_id)
+            .map_err(domain)?
+        {
+            // Published history must go through the ordinary continuity release.
+            return Ok(());
+        }
+        let operation_id = format!(
+            "subject-create-recovery:{}:{}",
+            lease.worker.worker_id, lease.generation
+        );
+        let fingerprint = api.config_store.prepare_subject_create_recovery(
+            api.workspace_id(),
+            &lease,
+            &operation_id,
+        )?;
+        let request = runtime_api::SubjectCreateRecoveryRequest {
+            workspace_id: api.workspace_id().into(),
+            worker_id: worker_runtime::identity::WorkerId::parse(&lease.worker.worker_id)
+                .ok_or_else(|| Error::Store("invalid retained Worker ID".into()))?,
+            operation_id,
+            create_fingerprint: fingerprint,
+        };
+        (lease, request)
+    };
+    // Never hold the callback lane/gate across Runtime I/O. Unknown outcome
+    // leaves the durable intent in place and reuses the exact identity next time.
+    let response = api
+        .runtime
+        .recover_subject_create(&lease.worker.runtime_id, request.clone())
+        .map_err(RuntimeRegistryError::into_error)?;
+    if response.workspace_id != request.workspace_id
+        || response.worker_id != request.worker_id
+        || response.operation_id != request.operation_id
+        || response.create_fingerprint != request.create_fingerprint
+    {
+        return Err(Error::RepositoryConflict(
+            "Subject create recovery acknowledgment mismatch; retain occupancy".into(),
+        ));
+    }
+    let _guard = gate
+        .lock()
+        .map_err(|_| Error::Store("Journal lifecycle lock poisoned".into()))?;
+    if !store
+        .journal_body_is_unpublished(subject, &lease.worker.runtime_id, &lease.worker.worker_id)
+        .map_err(domain)?
+    {
+        return Err(Error::RepositoryConflict(
+            "Subject publication appeared during recovery; retain occupancy".into(),
+        ));
+    }
+    api.config_store.complete_subject_create_recovery(
+        api.workspace_id(),
+        &lease,
+        &request.operation_id,
+        &request.create_fingerprint,
+    )
+}
+
 pub(super) fn lifecycle_lock(api: &WorkspaceApi, subject: &str) -> Result<Arc<Mutex<()>>> {
     let mut locks = api
         .journal_lifecycle

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

@@ -197,7 +197,20 @@ async fn rejected_runtime_proofs_log_reason_and_verification_without_secrets_on_
                 .unwrap();
             assert_eq!(response.status(), StatusCode::UNAUTHORIZED, "{reason}");
             let response_body = to_bytes(response.into_body(), usize::MAX).await.unwrap();
-            assert!(response_body.is_empty(), "{response_body:?}");
+            let error: server_api::RepositoryApiError =
+                serde_json::from_slice(&response_body).unwrap();
+            assert_eq!(error.error, "Unauthorized");
+            assert_eq!(error.message, "invalid runtime request proof");
+            assert_eq!(
+                serde_json::from_slice::<Value>(&response_body).unwrap(),
+                serde_json::to_value(server_api::RepositoryApiError::new(
+                    401,
+                    "Unauthorized",
+                    "invalid runtime request proof",
+                    Vec::new()
+                ))
+                .unwrap()
+            );
             let (event, text) = logs.event();
             assert_eq!(event["reason"], reason);
             assert_eq!(event["proof_verification"], verification);
@@ -338,6 +351,124 @@ async fn runtime_proof_setup_failure_logs_no_internal_error_and_bounds_request_m
     }
 }
 
+#[tokio::test]
+async fn common_api_rejections_use_json_without_changing_drive_or_non_api_contracts() {
+    for status in [
+        StatusCode::BAD_REQUEST,
+        StatusCode::UNAUTHORIZED,
+        StatusCode::FORBIDDEN,
+        StatusCode::NOT_FOUND,
+        StatusCode::CONFLICT,
+        StatusCode::SERVICE_UNAVAILABLE,
+    ] {
+        for path in [
+            "/api/w/ws/subjektiv/sessions",
+            "/api/w/ws/workers/self/workdir-session/operations",
+            "/api/runtime/v1/workspaces/ws/resources/fetch",
+            "/api/repositories",
+        ] {
+            let response = repository_api_rejection(path, status, "request rejected");
+            assert_eq!(response.status(), status);
+            assert_eq!(response.headers()[CONTENT_TYPE], "application/json");
+            let bytes = to_bytes(response.into_body(), usize::MAX).await.unwrap();
+            assert_eq!(
+                serde_json::from_slice::<Value>(&bytes).unwrap(),
+                serde_json::to_value(server_api::RepositoryApiError::new(
+                    status.as_u16(),
+                    status.canonical_reason().unwrap(),
+                    "request rejected",
+                    Vec::new()
+                ))
+                .unwrap()
+            );
+        }
+        let drive = repository_api_rejection("/api/w/ws/drive/mutate", status, "request rejected");
+        assert_eq!(drive.status(), status);
+        assert_eq!(
+            serde_json::from_slice::<Value>(
+                &to_bytes(drive.into_body(), usize::MAX).await.unwrap()
+            )
+            .unwrap(),
+            serde_json::to_value(drive::failure(status.as_u16())).unwrap()
+        );
+        let shell = repository_api_rejection("/w/ws", status, "request rejected");
+        assert_eq!(shell.status(), status);
+        assert!(
+            to_bytes(shell.into_body(), usize::MAX)
+                .await
+                .unwrap()
+                .is_empty()
+        );
+    }
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn subjektiv_auth_rejection_survives_runtime_generated_client_on_both_routers() {
+    use worker::WorkspaceClient;
+    struct AbortServer(tokio::task::JoinHandle<()>);
+    impl Drop for AbortServer {
+        fn drop(&mut self) {
+            self.0.abort();
+        }
+    }
+
+    let workspace = tempfile::tempdir().unwrap();
+    let mut api = test_api(workspace.path()).await;
+    let trusted = RuntimeIdentityMaterial::generate("runtime-test").unwrap();
+    configure_runtime_request_auth(&mut api, &trusted, "runtime-test");
+    let untrusted = RuntimeIdentityMaterial::generate("runtime-test").unwrap();
+    for app in [
+        build_router(api.clone()),
+        workspace_server_router(WorkspaceServerApi::new_for_test(
+            api.config.clone(),
+            api.store.clone(),
+        )),
+    ] {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let url = format!("http://{}", listener.local_addr().unwrap());
+        let _server = AbortServer(tokio::spawn(async move {
+            axum::serve(listener, app).await.unwrap();
+        }));
+        // Exercise missing authentication, signature rejection and a valid proof
+        // whose Worker has no catalog membership/reservation.
+        for identity in [None, Some(&untrusted), Some(&trusted)] {
+            let client = worker_runtime::worker_source::RuntimeOwnedWorkspaceClient::new(
+                TEST_WORKSPACE_ID,
+                &url,
+                "runtime-test",
+                "unregistered-worker",
+            );
+            let client = match identity {
+                Some(identity) => client.with_runtime_request_source(identity, "server-test"),
+                None => client,
+            };
+            let response = tokio::task::spawn_blocking(move || {
+                client.execute_server_operation(worker::WorkspaceServerOperation::SubjektivRecordSession(
+                    server_api::SubjektivRecordSessionRequest {
+                        session_id: session_store::new_session_id().to_string(), create_if_missing: true,
+                    },
+                ))
+            }).await.unwrap().expect("an explicit HTTP rejection must remain a WorkspaceResponse, not an unknown transport outcome");
+            assert_eq!(response.status, 401);
+            let error: server_api::RepositoryApiError =
+                serde_json::from_str(&response.body).unwrap();
+            assert_eq!(error.error, "Unauthorized");
+            assert_eq!(
+                error.message,
+                if identity.is_some() {
+                    "invalid runtime request proof"
+                } else {
+                    "authentication required"
+                }
+            );
+            assert!(!response.body.contains("unregistered-worker"));
+            if let Some(identity) = identity {
+                assert!(!response.body.contains(&identity.private_key));
+            }
+        }
+    }
+}
+
 #[tokio::test]
 async fn repository_runtime_proof_401_keeps_generic_envelope_on_both_routers() {
     let workspace = tempfile::tempdir().unwrap();

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

@@ -330,7 +330,7 @@ async fn rejected_create_releases_key_but_retains_subject_occupancy_without_rele
         .reserve_worker_create(
             TEST_WORKSPACE_ID,
             EMBEDDED_RUNTIME_ID,
-            "rejected-create",
+            "manual:rejected-create",
             &"f".repeat(64),
             Some("subjektiv:rejected-create"),
             &memory,
@@ -378,5 +378,274 @@ async fn rejected_create_releases_key_but_retains_subject_occupancy_without_rele
                 .worker,
             worker,
         );
+        // Retained Subject occupancy is not a live create reservation. A new
+        // manual request must not send the terminal Worker's identity to Runtime.
+        let retry = api.config_store.reserve_worker_create(
+            TEST_WORKSPACE_ID,
+            EMBEDDED_RUNTIME_ID,
+            "manual:retry-rejected-create",
+            &"f".repeat(64),
+            Some("subjektiv:rejected-create"),
+            &memory,
+        );
+        assert!(
+            matches!(retry, Err(Error::RepositoryConflict(ref message))
+                if message.contains("worker_singleton_release_required")),
+            "terminal reservation must fail before Runtime dispatch: {retry:?}"
+        );
     }
 }
+
+async fn unpublished_recovery_fixture(
+    root: &std::path::Path,
+) -> (
+    WorkspaceApi,
+    Arc<DeterministicExecutionBackend>,
+    String,
+    crate::store::WorkerCreateReservation,
+) {
+    let (api, backend) = test_api_with_recording_backend(root).await;
+    let subject = open_subjektiv_store(&api)
+        .unwrap()
+        .create_subject(crate::subjektiv::SubjectRole::new("recovery-test").unwrap())
+        .unwrap();
+    let key = crate::subjektiv::subject_worker_singleton_key(&subject.id).unwrap();
+    let memory = api
+        .config_store
+        .get_workspace_memory_settings(TEST_WORKSPACE_ID)
+        .unwrap();
+    let old = api
+        .config_store
+        .reserve_worker_create(
+            TEST_WORKSPACE_ID,
+            EMBEDDED_RUNTIME_ID,
+            "manual:failed-subject",
+            &"f".repeat(64),
+            Some(&key),
+            &memory,
+        )
+        .unwrap();
+    api.config_store
+        .fail_worker_create_reservation(
+            TEST_WORKSPACE_ID,
+            EMBEDDED_RUNTIME_ID,
+            old.worker_id,
+            &old.create_fingerprint,
+        )
+        .unwrap();
+    (api, backend, subject.id, old)
+}
+
+fn recovery_spawn_request(subject: &str) -> WorkerSpawnRequest {
+    serde_json::from_value(json!({
+        "intent": WorkerSpawnIntent::WorkspaceCompanion,
+        "acceptance": WorkerSpawnAcceptanceRequirement::RunAccepted { expected_segments: 0 },
+        "profile": worker_runtime::catalog::ProfileSelector::Builtin("builtin:companion".into()),
+        "singleton_key": crate::subjektiv::subject_worker_singleton_key(subject).unwrap(),
+    }))
+    .unwrap()
+}
+
+#[tokio::test]
+async fn unpublished_create_recovery_retries_same_fence_then_creates_new_worker() {
+    let root = tempfile::tempdir().unwrap();
+    let (api, backend, subject, old) = unpublished_recovery_fixture(root.path()).await;
+    let old_worker = RuntimeWorkerRef::new(EMBEDDED_RUNTIME_ID, old.worker_id.to_string());
+    let lease = old.singleton.as_ref().unwrap();
+    let late = crate::subjektiv::SubjectSessionAttribution::new(
+        &subject,
+        EMBEDDED_RUNTIME_ID,
+        old.worker_id.to_string(),
+        session_store::new_session_id().to_string(),
+    )
+    .unwrap();
+    let late_session = late.session_id.clone();
+    // Models an already-authorized callback arriving while Runtime fencing is
+    // in flight. The Host intent must reject it BEFORE attribution insertion.
+    let callback_api = api.clone();
+    *backend.recovery_hook.lock().unwrap() = Some(Box::new(move || {
+        assert!(subjektiv_journal::bind_host_session(&callback_api, late).is_err());
+    }));
+    *backend.recovery_failure.lock().unwrap() = Some("unconfirmed execution absence".into());
+    assert!(
+        api.spawn_workspace_worker(EMBEDDED_RUNTIME_ID, recovery_spawn_request(&subject))
+            .is_err()
+    );
+    assert!(
+        api.config_store
+            .subject_release_pending(TEST_WORKSPACE_ID, &old_worker)
+            .unwrap()
+    );
+    assert_eq!(
+        api.config_store
+            .current_worker_singleton_owner(TEST_WORKSPACE_ID, &lease.key)
+            .unwrap(),
+        Some(lease.clone())
+    );
+    let store = open_subjektiv_store(&api).unwrap();
+    assert!(store.session_attribution(&late_session).unwrap().is_none());
+    assert!(
+        store
+            .journal_body_is_unpublished(&subject, EMBEDDED_RUNTIME_ID, &old_worker.worker_id)
+            .unwrap()
+    );
+
+    let replacement = api
+        .spawn_workspace_worker(EMBEDDED_RUNTIME_ID, recovery_spawn_request(&subject))
+        .unwrap();
+    assert_eq!(replacement.state, InternalWorkerOperationState::Accepted);
+    let replacement = replacement.worker.unwrap().worker;
+    assert_ne!(replacement, old_worker);
+    let current = api
+        .config_store
+        .current_worker_singleton_owner(TEST_WORKSPACE_ID, &lease.key)
+        .unwrap()
+        .unwrap();
+    assert_eq!(current.worker, replacement);
+    assert_eq!(current.generation, lease.generation + 1);
+    let requests = backend.recovery_requests.lock().unwrap();
+    assert_eq!(requests.len(), 2);
+    assert_eq!(
+        requests[0], requests[1],
+        "unknown result must reuse exact recovery identity"
+    );
+    // A stale completion cannot release the replacement generation.
+    api.config_store
+        .complete_subject_create_recovery(
+            TEST_WORKSPACE_ID,
+            lease,
+            &requests[0].operation_id,
+            &old.create_fingerprint,
+        )
+        .unwrap();
+    assert_eq!(
+        api.config_store
+            .current_worker_singleton_owner(TEST_WORKSPACE_ID, &lease.key)
+            .unwrap(),
+        Some(current)
+    );
+}
+
+#[tokio::test]
+async fn unpublished_create_recovery_refuses_any_attribution_or_orphan_writer() {
+    for orphan_writer in [false, true] {
+        let root = tempfile::tempdir().unwrap();
+        let (api, backend, subject, old) = unpublished_recovery_fixture(root.path()).await;
+        let store = open_subjektiv_store(&api).unwrap();
+        if orphan_writer {
+            store
+                .claim_journal_writer(&subject, EMBEDDED_RUNTIME_ID, &old.worker_id.to_string())
+                .unwrap();
+        } else {
+            store
+                .record_session_attribution(
+                    crate::subjektiv::SubjectSessionAttribution::new(
+                        &subject,
+                        EMBEDDED_RUNTIME_ID,
+                        old.worker_id.to_string(),
+                        session_store::new_session_id().to_string(),
+                    )
+                    .unwrap(),
+                )
+                .unwrap();
+        }
+        assert!(
+            api.spawn_workspace_worker(EMBEDDED_RUNTIME_ID, recovery_spawn_request(&subject))
+                .is_err()
+        );
+        assert!(backend.recovery_requests.lock().unwrap().is_empty());
+        assert_eq!(
+            api.config_store
+                .current_worker_singleton_owner(
+                    TEST_WORKSPACE_ID,
+                    &old.singleton.as_ref().unwrap().key
+                )
+                .unwrap(),
+            old.singleton
+        );
+    }
+}
+
+#[tokio::test]
+async fn unpublished_create_recovery_resumes_after_runtime_proof_before_host_commit() {
+    let root = tempfile::tempdir().unwrap();
+    let (api, backend, subject, old) = unpublished_recovery_fixture(root.path()).await;
+    let lease = old.singleton.as_ref().unwrap();
+    let op = format!(
+        "subject-create-recovery:{}:{}",
+        old.worker_id, lease.generation
+    );
+    let fingerprint = api
+        .config_store
+        .prepare_subject_create_recovery(TEST_WORKSPACE_ID, lease, &op)
+        .unwrap();
+    let request = runtime_api::SubjectCreateRecoveryRequest {
+        workspace_id: TEST_WORKSPACE_ID.into(),
+        worker_id: old.worker_id,
+        create_fingerprint: fingerprint,
+        operation_id: op,
+    };
+    api.runtime
+        .recover_subject_create(EMBEDDED_RUNTIME_ID, request)
+        .unwrap();
+    // Simulate loss of the receipt before Host completion. The Host must not
+    // infer completion from missing catalog state, and can retry the exact op.
+    assert_eq!(
+        api.config_store
+            .current_worker_singleton_owner(TEST_WORKSPACE_ID, &lease.key)
+            .unwrap(),
+        old.singleton
+    );
+    subjektiv_journal::recover_failed_subject_create(&api, &subject).unwrap();
+    assert!(
+        api.config_store
+            .current_worker_singleton_owner(TEST_WORKSPACE_ID, &lease.key)
+            .unwrap()
+            .is_none()
+    );
+    assert_eq!(backend.recovery_requests.lock().unwrap().len(), 2);
+}
+
+#[tokio::test]
+async fn failed_subject_spawn_can_recover_on_next_explicit_create() {
+    let root = tempfile::tempdir().unwrap();
+    let (api, backend) = test_api_with_recording_backend(root.path()).await;
+    let store = open_subjektiv_store(&api).unwrap();
+    let subject = store
+        .create_subject(crate::subjektiv::SubjectRole::new("failed-spawn").unwrap())
+        .unwrap();
+    let key = crate::subjektiv::subject_worker_singleton_key(&subject.id).unwrap();
+    *backend.spawn_failure.lock().unwrap() =
+        Some("Session registration failed before publication".into());
+    let rejected = api
+        .spawn_workspace_worker(EMBEDDED_RUNTIME_ID, recovery_spawn_request(&subject.id))
+        .unwrap();
+    assert_eq!(rejected.state, InternalWorkerOperationState::Rejected);
+    let old = api
+        .config_store
+        .current_worker_singleton_owner(TEST_WORKSPACE_ID, &key)
+        .unwrap()
+        .unwrap();
+    assert!(
+        api.config_store
+            .subject_create_recovery_candidate(TEST_WORKSPACE_ID, &old)
+            .unwrap()
+            .is_some()
+    );
+    assert!(
+        backend.recovery_requests.lock().unwrap().is_empty(),
+        "failure alone must not release occupancy"
+    );
+    let result = api
+        .spawn_workspace_worker(EMBEDDED_RUNTIME_ID, recovery_spawn_request(&subject.id))
+        .unwrap();
+    assert_eq!(result.state, InternalWorkerOperationState::Accepted);
+    let current = api
+        .config_store
+        .current_worker_singleton_owner(TEST_WORKSPACE_ID, &key)
+        .unwrap()
+        .unwrap();
+    assert_ne!(current.worker, old.worker);
+    assert_eq!(current.generation, old.generation + 1);
+    assert_eq!(backend.recovery_requests.lock().unwrap().len(), 1);
+}

crates/workspace-server/src/store.rs

@@ -119,7 +119,7 @@ pub use ticket::{
 };
 
 const OLDEST_SCHEMA_VERSION: i64 = 50;
-const LATEST_SCHEMA_VERSION: i64 = 101;
+const LATEST_SCHEMA_VERSION: i64 = 102;
 const WORKSPACE_DRIVE_GRANTS_MIGRATION_NAME: &str = "durable Workspace Drive grants";
 const WORKSPACE_CONFIG_GRANTS_MIGRATION_NAME: &str = "Workspace config grants and logical Workdirs";
 const SCHEMA_BASELINE_NAME: &str = "workspace schema baseline";
@@ -432,6 +432,11 @@ const MIGRATIONS: &[Migration] = &[
         name: "Durable explicit Subject body release",
         apply: migrate_subject_release_v100_to_v101,
     },
+    Migration {
+        version: 102,
+        name: "Durable failed Subject creation recovery",
+        apply: migrate_subject_create_recovery_v101_to_v102,
+    },
 ];
 
 #[derive(Clone, Copy)]
@@ -2997,7 +3002,7 @@ impl SqliteWorkspaceStore {
                     .query_row(
                         "SELECT allocation_key, request_fingerprint, create_fingerprint, \
                                 memory_language, singleton_key, \
-                                singleton_generation \
+                                singleton_generation, state \
                          FROM worker_create_reservations \
                          WHERE workspace_id = ?1 AND runtime_id = ?2 AND worker_id = ?3",
                         params![
@@ -3013,11 +3018,21 @@ impl SqliteWorkspaceStore {
                                 row.get::<_, Option<String>>(3)?,
                                 row.get::<_, Option<String>>(4)?,
                                 row.get::<_, Option<i64>>(5)?,
+                                row.get::<_, String>(6)?,
                             ))
                         },
                     )
                     .optional()?;
-                if let Some((stored_allocation_key, stored_request_fingerprint, create_fingerprint, language, stored_singleton_key, stored_generation)) = stored
+                // Subject occupancy survives removal until an explicit release
+                // proof. It must not resurrect the removed creation reservation:
+                // Runtime callbacks require a live reservation or catalog member.
+                if stored.as_ref().is_some_and(|row| row.6 == "removed") {
+                    return Err(Error::RepositoryConflict(format!(
+                        "worker_singleton_release_required: Worker {}:{} has a terminal create reservation but still owns the Subject; complete Subject release before creating a replacement",
+                        owner.worker.runtime_id, owner.worker.worker_id
+                    )));
+                }
+                if let Some((stored_allocation_key, stored_request_fingerprint, create_fingerprint, language, stored_singleton_key, stored_generation, _)) = stored
                     && allocation_key.starts_with("manual:")
                     && stored_allocation_key.starts_with("manual:")
                     && owner.worker.runtime_id == runtime_id
@@ -3199,6 +3214,99 @@ impl SqliteWorkspaceStore {
         })
     }
 
+    /// Eligibility is Backend evidence only. The Host must separately prove that Subjektiv
+    /// never published this body and obtain the Runtime's exact create-fence receipt.
+    pub(crate) fn subject_create_recovery_candidate(
+        &self,
+        workspace: &str,
+        lease: &WorkerSingletonLease,
+    ) -> Result<Option<String>> {
+        self.with_conn(|conn| subject_create_recovery_candidate(conn, workspace, lease))
+    }
+
+    pub(crate) fn prepare_subject_create_recovery(
+        &self,
+        workspace: &str,
+        lease: &WorkerSingletonLease,
+        operation_id: &str,
+    ) -> Result<String> {
+        if operation_id.trim().is_empty() || operation_id.len() > 512 {
+            return Err(Error::InvalidInput(
+                "invalid Subject create recovery operation".into(),
+            ));
+        }
+        self.with_conn_mut(|conn| {
+            let tx = conn.transaction_with_behavior(TransactionBehavior::Immediate)?;
+            let fingerprint = subject_create_recovery_candidate(&tx, workspace, lease)?
+                .ok_or_else(|| Error::RepositoryConflict("Subject create recovery requires the current removed reservation without a registry row".into()))?;
+            let saved: Option<(String, String, String, u64, String)> = tx.query_row(
+                "SELECT runtime_id,worker_id,singleton_key,singleton_generation,create_fingerprint FROM subject_create_recoveries WHERE workspace_id=?1 AND operation_id=?2",
+                params![workspace, operation_id],
+                |row| Ok((row.get(0)?, row.get(1)?, row.get(2)?, row.get(3)?, row.get(4)?)),
+            ).optional()?;
+            if let Some(saved) = saved {
+                if saved != (lease.worker.runtime_id.clone(), lease.worker.worker_id.clone(), lease.key.clone(), lease.generation, fingerprint.clone()) {
+                    return Err(Error::RepositoryConflict("Subject create recovery operation was used with different input".into()));
+                }
+                return Ok(fingerprint);
+            }
+            let busy: bool = tx.query_row(
+                "SELECT EXISTS(SELECT 1 FROM subject_create_recoveries WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND singleton_key=?4 AND singleton_generation=?5)",
+                params![workspace, lease.worker.runtime_id, lease.worker.worker_id, lease.key, lease.generation as i64],
+                |row| row.get(0),
+            )?;
+            if busy {
+                return Err(Error::RepositoryConflict("Subject create recovery already has a durable operation".into()));
+            }
+            tx.execute(
+                "INSERT INTO subject_create_recoveries(workspace_id,runtime_id,worker_id,singleton_key,singleton_generation,operation_id,create_fingerprint) VALUES(?1,?2,?3,?4,?5,?6,?7)",
+                params![workspace, lease.worker.runtime_id, lease.worker.worker_id, lease.key, lease.generation as i64, operation_id, fingerprint],
+            )?;
+            tx.commit()?;
+            Ok(fingerprint)
+        })
+    }
+
+    /// Called only after the Host verifies the Runtime's typed receipt for this exact create.
+    /// The completed bit records that proof; it neither creates a writer binding nor rewrites
+    /// the reservation. An exact completed replay is a receipt, not authority over a successor.
+    pub(crate) fn complete_subject_create_recovery(
+        &self,
+        workspace: &str,
+        lease: &WorkerSingletonLease,
+        operation_id: &str,
+        create_fingerprint: &str,
+    ) -> Result<()> {
+        let generation = i64::try_from(lease.generation)
+            .map_err(|_| Error::InvalidInput("invalid singleton generation".into()))?;
+        self.with_conn_mut(|conn| {
+            let tx = conn.transaction_with_behavior(TransactionBehavior::Immediate)?;
+            let completed: Option<bool> = tx.query_row(
+                "SELECT completed FROM subject_create_recoveries WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND singleton_key=?4 AND singleton_generation=?5 AND operation_id=?6 AND create_fingerprint=?7",
+                params![workspace, lease.worker.runtime_id, lease.worker.worker_id, lease.key, generation, operation_id, create_fingerprint],
+                |row| row.get(0),
+            ).optional()?;
+            match completed {
+                Some(true) => return Ok(()),
+                Some(false) => {},
+                None => return Err(Error::RepositoryConflict("Subject create recovery intent absent or mismatched".into())),
+            }
+            let current = subject_create_recovery_candidate(&tx, workspace, lease)?;
+            if current.as_deref() != Some(create_fingerprint) {
+                return Err(Error::RepositoryConflict("Subject create recovery reservation or singleton ownership changed".into()));
+            }
+            let changed = tx.execute(
+                "UPDATE subject_create_recoveries SET completed=1 WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND singleton_key=?4 AND singleton_generation=?5 AND operation_id=?6 AND create_fingerprint=?7 AND completed=0",
+                params![workspace, lease.worker.runtime_id, lease.worker.worker_id, lease.key, generation, operation_id, create_fingerprint],
+            )?;
+            if changed != 1 {
+                return Err(Error::RepositoryConflict("Subject create recovery completion lost its exact intent".into()));
+            }
+            tx.commit()?;
+            Ok(())
+        })
+    }
+
     pub(crate) fn retained_subject_release_operation(
         &self,
         workspace: &str,
@@ -3213,7 +3321,7 @@ impl SqliteWorkspaceStore {
         workspace: &str,
         worker: &RuntimeWorkerRef,
     ) -> Result<bool> {
-        self.with_conn(|c| Ok(c.query_row("SELECT EXISTS(SELECT 1 FROM subject_worker_releases WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND cancelled=0 AND completed=0)",params![workspace,worker.runtime_id,worker.worker_id],|r|r.get(0))?))
+        self.with_conn(|c| Ok(c.query_row("SELECT EXISTS(SELECT 1 FROM subject_worker_releases WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND cancelled=0 AND completed=0) OR EXISTS(SELECT 1 FROM subject_create_recoveries WHERE workspace_id=?1 AND runtime_id=?2 AND worker_id=?3 AND completed=0)",params![workspace,worker.runtime_id,worker.worker_id],|r|r.get(0))?))
     }
 
     pub(crate) fn require_subject_release_intent(
@@ -15012,6 +15120,34 @@ fn migrate_result_reports_v97_to_v98(conn: &Connection) -> Result<()> {
     Ok(())
 }
 
+fn migrate_subject_create_recovery_v101_to_v102(conn: &Connection) -> Result<()> {
+    if current_schema_version(conn)? != 101 {
+        return Err(Error::Store(
+            "Subject create recovery migration requires schema 101".into(),
+        ));
+    }
+    let tx = conn.unchecked_transaction()?;
+    tx.execute_batch("CREATE TABLE subject_create_recoveries (
+    workspace_id TEXT NOT NULL,
+    runtime_id TEXT NOT NULL,
+    worker_id TEXT NOT NULL,
+    singleton_key TEXT NOT NULL,
+    singleton_generation INTEGER NOT NULL CHECK (singleton_generation > 0),
+    operation_id TEXT NOT NULL CHECK (length(operation_id) > 0),
+    create_fingerprint TEXT NOT NULL,
+    completed INTEGER NOT NULL DEFAULT 0 CHECK (completed IN (0,1)),
+    PRIMARY KEY (workspace_id, operation_id),
+    UNIQUE (workspace_id, runtime_id, worker_id, singleton_key, singleton_generation),
+    FOREIGN KEY (workspace_id, worker_id) REFERENCES worker_create_reservations(workspace_id, worker_id) ON DELETE CASCADE
+);")?;
+    tx.execute(
+        "INSERT INTO __yoi_schema_migrations(version,name) VALUES(102,?1)",
+        ["Durable failed Subject creation recovery"],
+    )?;
+    tx.commit()?;
+    Ok(())
+}
+
 fn migrate_subject_release_v100_to_v101(conn: &Connection) -> Result<()> {
     if current_schema_version(conn)? != 100 {
         return Err(Error::Store(
@@ -15702,6 +15838,33 @@ pub(crate) fn validate_worker_singleton_key(key: &str) -> Result<()> {
     Ok(())
 }
 
+fn subject_create_recovery_candidate(
+    conn: &Connection,
+    workspace: &str,
+    lease: &WorkerSingletonLease,
+) -> Result<Option<String>> {
+    if !lease
+        .key
+        .starts_with(crate::subjektiv::SUBJECT_WORKER_SINGLETON_PREFIX)
+    {
+        return Err(Error::InvalidInput("not a Subject singleton".into()));
+    }
+    let generation = i64::try_from(lease.generation)
+        .map_err(|_| Error::InvalidInput("invalid singleton generation".into()))?;
+    Ok(conn.query_row(
+        "SELECT r.create_fingerprint FROM worker_create_reservations r
+         JOIN worker_singleton_owners o ON o.workspace_id=r.workspace_id
+           AND o.runtime_id=r.runtime_id AND o.worker_id=r.worker_id
+           AND o.singleton_key=r.singleton_key AND o.generation=r.singleton_generation
+         WHERE r.workspace_id=?1 AND r.runtime_id=?2 AND r.worker_id=?3
+           AND r.singleton_key=?4 AND r.singleton_generation=?5 AND r.state='removed'
+           AND NOT EXISTS(SELECT 1 FROM worker_registry w
+             WHERE w.workspace_id=r.workspace_id AND w.runtime_id=r.runtime_id AND w.worker_id=r.worker_id)",
+        params![workspace, lease.worker.runtime_id, lease.worker.worker_id, lease.key, generation],
+        |row| row.get(0),
+    ).optional()?)
+}
+
 fn current_worker_singleton_owner(
     conn: &Connection,
     workspace_id: &str,
@@ -15712,6 +15875,7 @@ fn current_worker_singleton_owner(
             "SELECT owner.runtime_id, owner.worker_id, owner.generation, \
                     CASE WHEN substr(owner.singleton_key,1,10)='subjektiv:' AND reservation.state='removed' \
                       AND NOT EXISTS(SELECT 1 FROM subject_worker_releases r WHERE r.workspace_id=owner.workspace_id AND r.worker_id=owner.worker_id AND r.runtime_id=owner.runtime_id AND r.singleton_key=owner.singleton_key AND r.singleton_generation=owner.generation AND r.completed=1) \
+                      AND NOT EXISTS(SELECT 1 FROM subject_create_recoveries r WHERE r.workspace_id=owner.workspace_id AND r.worker_id=owner.worker_id AND r.runtime_id=owner.runtime_id AND r.singleton_key=owner.singleton_key AND r.singleton_generation=owner.generation AND r.create_fingerprint=reservation.create_fingerprint AND r.completed=1) \
                     THEN 'created' ELSE reservation.state END \
              FROM worker_singleton_owners owner \
              LEFT JOIN worker_create_reservations reservation \
@@ -18739,6 +18903,7 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
                 WorkspaceSchemaMigrationStep {version:99, name:"Explicit hosting administrators and self-registration policy".into()},
                 WorkspaceSchemaMigrationStep {version:100, name:"Provisional users and scoped passkey activation".into()},
                 WorkspaceSchemaMigrationStep {version:101, name:"Durable explicit Subject body release".into()},
+                WorkspaceSchemaMigrationStep {version:102, name:"Durable failed Subject creation recovery".into()},
             ]
         );
 
@@ -18850,6 +19015,7 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
 (99, "Explicit hosting administrators and self-registration policy".to_string()),
 (100, "Provisional users and scoped passkey activation".to_string()),
 (101, "Durable explicit Subject body release".to_string()),
+(102, "Durable failed Subject creation recovery".to_string()),
                     ]
                 );
                 assert!(!table_exists(conn, "trusted_runtime_records")?);
@@ -20907,6 +21073,47 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
             conflict,
             Error::RepositoryConflict(message) if message.contains("worker_singleton_owned")
         ));
+
+        reopened
+            .fail_worker_create_reservation(
+                "workspace-a",
+                "runtime-a",
+                reserved.worker_id,
+                &reserved.create_fingerprint,
+            )
+            .unwrap();
+        drop(reopened);
+        let reopened = SqliteWorkspaceStore::open_for_test(&database).unwrap();
+        let conflict = reopened
+            .reserve_worker_create(
+                "workspace-a",
+                "runtime-a",
+                "manual:retry-after-terminal-failure",
+                &stable_fingerprint,
+                Some("subjektiv:restart-retry"),
+                &memory,
+            )
+            .unwrap_err();
+        assert!(matches!(conflict, Error::RepositoryConflict(message)
+            if message.contains("worker_singleton_release_required")));
+        let worker = RuntimeWorkerRef::new("runtime-a", reserved.worker_id.to_string());
+        assert!(
+            !reopened
+                .has_active_worker_create_reservation("workspace-a", &worker)
+                .unwrap()
+        );
+        assert!(
+            reopened
+                .worker_resource_key("workspace-a", &worker)
+                .unwrap()
+                .is_none()
+        );
+        assert_eq!(
+            reopened
+                .current_worker_singleton_owner("workspace-a", "subjektiv:restart-retry")
+                .unwrap(),
+            reserved.singleton
+        );
     }
 
     #[tokio::test]
@@ -21036,6 +21243,383 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
         );
     }
 
+    #[test]
+    fn subject_create_recovery_migration_is_atomic_and_matches_fresh_schema() {
+        let dir = tempfile::tempdir().unwrap();
+        let database = dir.path().join("server.db");
+        prepare_retained_schema(&database, 101);
+        let conn = Connection::open(&database).unwrap();
+        configure_sqlite(&conn).unwrap();
+        conn.execute_batch("CREATE TRIGGER interrupt_create_recovery BEFORE INSERT ON __yoi_schema_migrations WHEN NEW.version=102 BEGIN SELECT RAISE(ABORT,'injected recovery migration interruption'); END;").unwrap();
+        assert!(migrate_subject_create_recovery_v101_to_v102(&conn).is_err());
+        assert_eq!(current_schema_version(&conn).unwrap(), 101);
+        assert!(!table_exists(&conn, "subject_create_recoveries").unwrap());
+        conn.execute_batch("DROP TRIGGER interrupt_create_recovery")
+            .unwrap();
+        migrate_subject_create_recovery_v101_to_v102(&conn).unwrap();
+        assert_eq!(current_schema_version(&conn).unwrap(), 102);
+        assert!(migrate_subject_create_recovery_v101_to_v102(&conn).is_err());
+        let fresh = SqliteWorkspaceStore::in_memory_for_test().unwrap();
+        let schema = |c: &Connection| -> Result<String> {
+            Ok(c.query_row(
+                "SELECT sql FROM sqlite_schema WHERE name='subject_create_recoveries'",
+                [],
+                |r| r.get::<_, String>(0),
+            )?
+            .split_whitespace()
+            .collect::<Vec<_>>()
+            .join(" "))
+        };
+        assert_eq!(schema(&conn).unwrap(), fresh.with_conn(schema).unwrap());
+        assert_eq!(
+            conn.query_row("SELECT count(*) FROM pragma_foreign_key_check", [], |r| r
+                .get::<_, i64>(
+                0
+            ))
+            .unwrap(),
+            0
+        );
+        drop(conn);
+        let reopened = SqliteWorkspaceStore::open_for_test(&database).unwrap();
+        reopened
+            .with_conn(|c| verify_schema_history(c, 102))
+            .unwrap();
+    }
+
+    #[tokio::test]
+    async fn subject_create_recovery_is_durable_exact_and_cannot_release_a_new_generation() {
+        for baseline in [None, Some(52), Some(101)] {
+            let dir = tempfile::tempdir().unwrap();
+            let database = dir.path().join("server.db");
+            if let Some(baseline) = baseline {
+                prepare_retained_schema(&database, baseline);
+            }
+            let store = SqliteWorkspaceStore::open_api_fixture(&database).unwrap();
+            store
+                .upsert_workspace(&WorkspaceRecord {
+                    workspace_id: "workspace-a".into(),
+                    owner_account_id: "owner-account".into(),
+                    display_name: "Workspace A".into(),
+                    state: "active".into(),
+                    created_at: "2026-10-01T00:00:00Z".into(),
+                    updated_at: "2026-10-01T00:00:00Z".into(),
+                })
+                .await
+                .unwrap();
+            let memory = store.get_workspace_memory_settings("workspace-a").unwrap();
+            let key = "subjektiv:failed-create";
+            let first = store
+                .reserve_worker_create(
+                    "workspace-a",
+                    "runtime-a",
+                    "manual:first",
+                    "sha256:first",
+                    Some(key),
+                    &memory,
+                )
+                .unwrap();
+            let lease = first.singleton.unwrap();
+            let fingerprint = first.create_fingerprint;
+            assert_eq!(
+                store
+                    .subject_create_recovery_candidate("workspace-a", &lease)
+                    .unwrap(),
+                None
+            );
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "recover-first")
+                    .is_err()
+            );
+            store
+                .fail_worker_create_reservation(
+                    "workspace-a",
+                    "runtime-a",
+                    first.worker_id,
+                    &fingerprint,
+                )
+                .unwrap();
+            assert_eq!(
+                store
+                    .subject_create_recovery_candidate("workspace-a", &lease)
+                    .unwrap(),
+                Some(fingerprint.clone())
+            );
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-b", &lease, "recover-first")
+                    .is_err()
+            );
+            for wrong in [
+                WorkerSingletonLease {
+                    generation: lease.generation + 1,
+                    ..lease.clone()
+                },
+                WorkerSingletonLease {
+                    key: "subjektiv:other".into(),
+                    ..lease.clone()
+                },
+                WorkerSingletonLease {
+                    worker: RuntimeWorkerRef::new("runtime-b", &lease.worker.worker_id),
+                    ..lease.clone()
+                },
+                WorkerSingletonLease {
+                    worker: RuntimeWorkerRef::new("runtime-a", "other-worker"),
+                    ..lease.clone()
+                },
+            ] {
+                assert_eq!(
+                    store
+                        .subject_create_recovery_candidate("workspace-a", &wrong)
+                        .unwrap(),
+                    None
+                );
+                assert!(
+                    store
+                        .prepare_subject_create_recovery("workspace-a", &wrong, "recover-first")
+                        .is_err()
+                );
+            }
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &lease,
+                        "recover-first",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            assert_eq!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "recover-first")
+                    .unwrap(),
+                fingerprint
+            );
+            assert!(
+                store
+                    .subject_release_pending("workspace-a", &lease.worker)
+                    .unwrap()
+            );
+            assert_eq!(
+                store
+                    .current_worker_singleton_owner("workspace-a", key)
+                    .unwrap(),
+                Some(lease.clone())
+            );
+            assert!(
+                store
+                    .reserve_worker_create(
+                        "workspace-a",
+                        "runtime-a",
+                        "manual:retry",
+                        "sha256:first",
+                        Some(key),
+                        &memory
+                    )
+                    .is_err()
+            );
+            drop(store);
+            let store = SqliteWorkspaceStore::open_for_test(&database).unwrap();
+            assert_eq!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "recover-first")
+                    .unwrap(),
+                fingerprint
+            );
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "different-operation")
+                    .is_err()
+            );
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &lease,
+                        "recover-first",
+                        "different-fingerprint"
+                    )
+                    .is_err()
+            );
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &lease,
+                        "different-operation",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-b",
+                        &lease,
+                        "recover-first",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            // A registry row appearing after prepare must block both replay and completion.
+            store.with_conn(|c| {
+                c.execute("INSERT INTO worker_registry(workspace_id,runtime_id,worker_id,display_name,retention_state,created_at,updated_at) VALUES('workspace-a',?1,?2,'body','normal','now','now')", params![lease.worker.runtime_id, lease.worker.worker_id])?;
+                Ok(())
+            }).unwrap();
+            assert_eq!(
+                store
+                    .subject_create_recovery_candidate("workspace-a", &lease)
+                    .unwrap(),
+                None
+            );
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "recover-first")
+                    .is_err()
+            );
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &lease,
+                        "recover-first",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            store.with_conn(|c| {
+                c.execute("DELETE FROM worker_registry WHERE workspace_id='workspace-a' AND runtime_id=?1 AND worker_id=?2", params![lease.worker.runtime_id, lease.worker.worker_id])?;
+                // Exercise a current-generation change while the operation is incomplete.
+                c.execute("UPDATE worker_singleton_owners SET generation=generation+1 WHERE workspace_id='workspace-a' AND singleton_key=?1", [key])?;
+                Ok(())
+            }).unwrap();
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &lease,
+                        "recover-first",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            assert!(
+                store
+                    .subject_release_pending("workspace-a", &lease.worker)
+                    .unwrap()
+            );
+            store.with_conn(|c| {
+                c.execute("UPDATE worker_singleton_owners SET generation=?2 WHERE workspace_id='workspace-a' AND singleton_key=?1", params![key, lease.generation as i64])?;
+                Ok(())
+            }).unwrap();
+            store
+                .complete_subject_create_recovery(
+                    "workspace-a",
+                    &lease,
+                    "recover-first",
+                    &fingerprint,
+                )
+                .unwrap();
+            drop(store);
+            let store = SqliteWorkspaceStore::open_for_test(&database).unwrap();
+            store
+                .complete_subject_create_recovery(
+                    "workspace-a",
+                    &lease,
+                    "recover-first",
+                    &fingerprint,
+                )
+                .unwrap();
+            assert!(
+                !store
+                    .subject_release_pending("workspace-a", &lease.worker)
+                    .unwrap()
+            );
+            assert_eq!(
+                store
+                    .current_worker_singleton_owner("workspace-a", key)
+                    .unwrap(),
+                None
+            );
+            let next = store
+                .reserve_worker_create(
+                    "workspace-a",
+                    "runtime-b",
+                    "manual:next",
+                    "sha256:next",
+                    Some(key),
+                    &memory,
+                )
+                .unwrap();
+            let next_lease = next.singleton.unwrap();
+            assert_eq!(next_lease.generation, lease.generation + 1);
+            store
+                .complete_subject_create_recovery(
+                    "workspace-a",
+                    &lease,
+                    "recover-first",
+                    &fingerprint,
+                )
+                .unwrap();
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &lease, "recover-first")
+                    .is_err()
+            );
+            assert!(
+                store
+                    .complete_subject_create_recovery(
+                        "workspace-a",
+                        &next_lease,
+                        "recover-first",
+                        &fingerprint
+                    )
+                    .is_err()
+            );
+            assert_eq!(
+                store
+                    .current_worker_singleton_owner("workspace-a", key)
+                    .unwrap(),
+                Some(next_lease.clone())
+            );
+            // The old completed proof cannot release the next removed generation.
+            store
+                .fail_worker_create_reservation(
+                    "workspace-a",
+                    "runtime-b",
+                    next.worker_id,
+                    &next.create_fingerprint,
+                )
+                .unwrap();
+            assert_eq!(
+                store
+                    .current_worker_singleton_owner("workspace-a", key)
+                    .unwrap(),
+                Some(next_lease.clone())
+            );
+            assert!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &next_lease, "recover-first")
+                    .is_err()
+            );
+            assert_eq!(
+                store
+                    .prepare_subject_create_recovery("workspace-a", &next_lease, "recover-next")
+                    .unwrap(),
+                next.create_fingerprint
+            );
+            store.with_conn(|c| {
+                assert_eq!(c.query_row("SELECT count(*) FROM subject_worker_releases", [], |r| r.get::<_, i64>(0))?, 0);
+                assert_eq!(c.query_row("SELECT state FROM worker_create_reservations WHERE workspace_id='workspace-a' AND worker_id=?1", [&lease.worker.worker_id], |r| r.get::<_, String>(0))?, "removed");
+                assert!(c.execute("UPDATE subject_create_recoveries SET completed=2", []).is_err());
+                assert!(c.execute("UPDATE subject_create_recoveries SET singleton_generation=0", []).is_err());
+                assert_eq!(c.query_row("SELECT count(*) FROM pragma_foreign_key_check", [], |r| r.get::<_, i64>(0))?, 0);
+                Ok(())
+            }).unwrap();
+        }
+    }
+
     #[test]
     fn subject_release_migration_rolls_back_schema_and_marker_then_reopens() {
         let dir = tempfile::tempdir().unwrap();
@@ -21056,7 +21640,7 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
         drop(conn);
         let reopened = SqliteWorkspaceStore::open_for_test(database).unwrap();
         reopened
-            .with_conn(|conn| verify_schema_history(conn, 101))
+            .with_conn(|conn| verify_schema_history(conn, LATEST_SCHEMA_VERSION))
             .unwrap();
     }
 
@@ -21636,7 +22220,8 @@ CREATE UNIQUE INDEX ticket_current_worker_role_idx
                          (98, "Assignment-independent implementation reports".into()),
                          (99, "Explicit hosting administrators and self-registration policy".into()),
                          (100, "Provisional users and scoped passkey activation".into()),
-                         (101, "Durable explicit Subject body release".into())]);
+                         (101, "Durable explicit Subject body release".into()),
+                         (102, "Durable failed Subject creation recovery".into())]);
                 for (table, rows) in unchanged.iter().zip(before.iter()) {
                     assert_eq!(&stored_fixture_rows(conn, table), rows, "preserve {table}");
                 }

crates/workspace-server/src/workspace_deletion.rs

@@ -95,6 +95,7 @@ const WORKSPACE_DELETION_PURGE_TABLES: &[&str] = &[
     "workdir_removal_operations",
     "worker_control_grants",
     "subject_worker_releases",
+    "subject_create_recoveries",
     "worker_create_reservations",
     "worker_diagnostics_archives",
     "worker_mutation_source_proof_jtis",