@memberjunction/server 6.2.0-edge.5 → 6.2.0-edge.7

package/dist/index.js

@@ -9,6 +9,8 @@ import { setupSQLServerClient, SQLServerDataProvider, SQLServerProviderConfigDat
 import { extendConnectionPoolWithQuery } from './util.js';
 import { registerIntegrationCustomColumnPromoter, IntegrationCustomColumnPromoter } from './integration/CustomColumnPromoter.js';
 import { DisableUnselectedEntityMaps, ReenableFieldMapsForEntityMap, selectFieldsToMap } from './integration/EntityMapLifecycle.js';
+import { RSUPostRestartProgressSession } from './integration/RSUPostRestartProgress.js';
+import { IntrospectInCatalogScope } from './integration/PostRestartIntrospect.js';
 import { default as BodyParser } from 'body-parser';
 import compression from 'compression'; // Add compression middleware
 import cors from 'cors';
@@ -29,13 +31,14 @@ import sql from 'mssql';
 import { WebSocketServer } from 'ws';
 import { RealtimeProxyServer } from './realtimeProxy/RealtimeProxyServer.js';
 import buildApolloServer from './apolloServer/index.js';
-import { configInfo, configFilePath, dbDatabase, dbHost, dbPort, dbUsername, graphqlPort, graphqlRootPath, mj_core_schema, websiteRunFromPackage } from './config.js';
+import { configInfo, configFilePath, dbDatabase, dbHost, dbPort, dbUsername, graphqlPort, graphqlRootPath, mj_core_schema, websiteRunFromPackage, userEmailMap } from './config.js';
+import { DescribeUserEmailMap } from './auth/loginEmailResolution.js';
 import { TranslateBracketsToPG } from './postgresqlCompat.js';
 import { BuildPostgreSQLConnectionConfig, DescribeReadOnlyLoginOverreach, PostgreSQLReadOnlyPool, ResolvePostgreSQLEndpoint, ResolvePostgreSQLReadOnlyCredentials, ToPGPoolConfig } from './postgresqlPoolSettings.js';
 import { default as jwt } from 'jsonwebtoken';
 import { contextFunction, CreateReadOnlyProvider, CreateUnifiedAuthMiddleware, getUserPayload } from './context.js';
 import { requireSystemUserDirective, publicDirective } from './directives/index.js';
-import { variablesLoggingMiddleware } from './logging/variablesLoggingMiddleware.js';
+import { ApplyRootVariablesLoggingMiddleware } from './logging/variablesLoggingMiddleware.js';
 import { auditResolversForUndecoratedArgs } from './logging/bootAudit.js';
 import { StartupLogger } from './logging/StartupLogger.js';
 import { AuthProviderFactory } from '@memberjunction/auth-providers';
@@ -43,7 +46,7 @@ import createMSSQLConfig from './orm.js';
 import { setupRESTEndpoints } from './rest/setupRESTEndpoints.js';
 import { CreateRealtimeRecordingRouter, REALTIME_RECORDING_MOUNT_PATH } from './rest/RealtimeRecordingHandler.js';
 import { LoadOwnedRealtimeSession } from './realtimeSessions/ownedRealtimeSession.js';
-import { LoadSessionConversationExternalID } from './realtimeSessions/sessionAccess.js';
+import { AuthorizeAnonymousRoomSessions, AuthorizeLoadedSession } from './realtimeSessions/sessionAccess.js';
 import { RecordingAssemblyScheduler } from './realtimeSessions/RecordingAssemblyScheduler.js';
 import { RealtimeSessionEventService, ParseReplicatedRealtimeSessionEvent, } from './realtimeSessions/index.js';
 import { resolve } from 'node:path';
@@ -51,20 +54,21 @@ import { DataSourceInfo, raiseEvent } from './types.js';
 import { ScheduledJobsService } from './services/ScheduledJobsService.js';
 import { IntegrationSyncWorkerService } from './services/IntegrationSyncWorkerService.js';
 import { LocalCacheManager, TelemetryManager, LogStatus, LogError, SetVerboseLogging } from '@memberjunction/core';
-import { validateAuthProvidersRegistered } from './auth/index.js';
+import { getSystemUser, validateAuthProvidersRegistered } from './auth/index.js';
 import { GetAPIKeyEngine } from '@memberjunction/api-keys';
 import { ActionEngineServer } from '@memberjunction/actions';
 import { CreateAdhocSQLAuthorizer } from './resolvers/AdhocQueryResolver.js';
 import { CacheManagerConfigFromSettings, CreateSharedCacheFromEnvironment, StartEngineSweeper, StartMetadataSweep, StartUserCacheChecks, WarmupLeaseMsFromSettings, WirePushStatusFanOut, WireSharedCacheEvents } from './sharedCache.js';
 import { PubSubManager } from './generic/PubSubManager.js';
-import { ReconcileOrphanedConversationDetails } from './generic/OrphanedConversationDetailReconciler.js';
+import { StartOrphanedConversationDetailReconciler } from './generic/OrphanedConversationDetailReconciler.js';
 import { HANDOFF_OFFER_FANOUT_CHANNEL, HANDOFF_OFFER_TOPIC, SetHandoffOfferPublishHook, ParseReplicatedHandoffOfferUpdate, } from './resolvers/HumanHandoffResolver.js';
-import { RoomHandoffEngine } from '@memberjunction/livekit-room-server';
-import { IntegrationProgressEmitter } from '@memberjunction/integration-progress-artifacts';
+import { RoomAuthorizationService, RoomHandoffEngine } from '@memberjunction/livekit-room-server';
+import { IntegrationProgressEmitter, SweepInterruptedRuns } from '@memberjunction/integration-progress-artifacts';
 import { PublishIntegrationProgress } from './resolvers/IntegrationProgressResolver.js';
 import { RegisterRSUProgressBridge } from './integration/RSUProgressBridge.js';
 import { ClientToolRequestManager, AgentRunWatchdog, SqlRealtimeRecordingLedger } from '@memberjunction/ai-agents';
 import { BindBridgeEngineHostInstance, SessionJanitor, SetHostInstancePort, StartBridgeOrphanReconciliation } from './agentSessions/index.js';
+import { ApplyHandoffOfferExpirySettings } from './agentSessions/HandoffOfferExpiry.js';
 import { StartTaskGraphDispatcher } from './services/StartTaskGraphDispatcher.js';
 import { MJServerWorkQueueProviderSource, StartWorkQueueHost } from './services/WorkQueueHostService.js';
 import { GetAttachmentService } from '@memberjunction/aiengine';
@@ -72,15 +76,22 @@ import { MJStorageBlobStore } from './services/MJStorageBlobStore.js';
 import { CACHE_INVALIDATION_TOPIC, MayBroadcastRecordData, ConfigureRecordDataBroadcast } from './generic/CacheInvalidationResolver.js';
 import { ConnectorFactory, IntegrationEngine } from '@memberjunction/integration-engine';
 import { CronExpressionHelper } from '@memberjunction/scheduling-engine';
-import { ServerExtensionLoader, mergeServerExtensionConfigs, prepareServerExtensionConfigs, describeServerExtensionMount, InstallMediaUpgradeDispatcher, IsGraphQLWsPath, GetCorsOwningServerExtensionRoots } from '@memberjunction/server-extensions-core';
+import { ServerExtensionLoader, SERVER_EXTENSION_CORE_SERVICE_KEYS, mergeServerExtensionConfigs, prepareServerExtensionConfigs, describeServerExtensionMount, InstallMediaUpgradeDispatcher, IsGraphQLWsPath, GetCorsOwningServerExtensionRoots, } from '@memberjunction/server-extensions-core';
+import { CreateIsolatedProvider } from './isolatedProvider.js';
+import { ResolveConfiguredPrincipal } from './auth/principals.js';
 import { CreateExtensionOwnedPreflightFallback, CreateGlobalCorsMiddleware } from './globalCors.js';
 import { CoreReservedServerExtensionRoots } from './serverExtensionReservedRoots.js';
 import { ApiRoute, ApiRootPath, ConfiguredPublicUrl } from './apiRoot.js';
 import { MountHealthcheckRoute, MountPublicCoreRoutes, MountExtensionHealthRoute, MountAuthCatalogRoute, MountAuthenticatedCoreRoutes } from './coreRoutes.js';
 import { MetadataCacheRefreshIntervalSeconds } from './providerConfigUnits.js';
 import { CreateMetadataRefreshSignalHandler, METADATA_REFRESH_SIGNAL } from './metadataRefreshSignal.js';
+import { CacheInvalidationBroadcastManager } from './broadcast/index.js';
+import { ResolveServerRole, ShouldRunBackgroundWorkers } from './serverRoleConfigUnits.js';
 const cacheRefreshInterval = configInfo.databaseSettings.metadataCacheRefreshInterval;
 export { MetadataCacheRefreshIntervalSeconds } from './providerConfigUnits.js';
+export * from './serverRoleConfigUnits.js';
+export * from './broadcast/index.js';
+export { ApplyRootVariablesLoggingMiddleware } from './logging/variablesLoggingMiddleware.js';
 /**
  * Returns the configured database platform from the `DB_PLATFORM` environment
  * variable, falling back to `'sqlserver'` when the env var is unset. An
@@ -115,6 +126,7 @@ export * from './directives/index.js';
 export { NoLog, hasNoLogParameter, getNoLogFields } from './logging/NoLog.js';
 export * from './entitySubclasses/MJEntityPermissionEntityServer.server.js';
 export * from './types.js';
+export * from './requestOrigin.js';
 export { getSystemUser, getSigningKeys, extractUserInfoFromPayload, verifyUserRecord, } from './auth/index.js';
 export * from './auth/APIKeyScopeAuth.js';
 export * from './auth/actingContextResolver.js';
@@ -263,8 +275,6 @@ function resolveServerVersion() {
         return undefined;
     }
 }
-/** How often to re-check for conversation details left behind by finished runs. */
-const ORPHAN_DETAIL_SWEEP_INTERVAL_MS = 5 * 60_000;
 /**
  * Replicate handoff-offer updates across server instances over Redis.
  *
@@ -359,6 +369,9 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     SetHostInstancePort(graphqlPort);
     // The realtime bridge engine stamps its bridge rows with the same identity and reconciles orphans by it (#5310).
     BindBridgeEngineHostInstance();
+    // An anonymous visitor may join only the LiveKit agent-session room it owns by scope; the room gate cannot import
+    // this package's session-access decision, so it is handed in here, before any resolver runs (MJ#6044).
+    RoomAuthorizationService.Instance.SetAnonymousSessionOwnerCheck(AuthorizeAnonymousRoomSessions);
     // Level-gated startup logger. Resolves verbosity from telemetry.level (single
     // operator knob). At `standard` (default), per-phase timings are collapsed into
     // the one-line summary; at `verbose`+ each phase prints inline as before.
@@ -411,6 +424,8 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
         // Every API pool carries the statement and idle-in-transaction timeouts from connection #1.
         const pgConnectionConfig = BuildPostgreSQLConnectionConfig(pgEndpoint, configInfo.databaseSettings, 'api');
         const pgPool = new pg.default.Pool(ToPGPoolConfig(pgConnectionConfig));
+        // An idle connection the server drops is emitted on the pool; without a listener Node exits.
+        pgPool.on('error', (err) => console.error('[PostgreSQL pool] Idle connection error (connection evicted):', err.message));
         // Verify connection
         const testClient = await pgPool.connect();
         await testClient.query('SELECT 1');
@@ -426,6 +441,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                 ...ToPGPoolConfig(BuildPostgreSQLConnectionConfig({ ...pgEndpoint, ...pgReadOnlyCredentials }, configInfo.databaseSettings, 'read-only')),
                 types: MJPostgresTypes,
             });
+            readOnlyPgPool.on('error', (err) => console.error('[PostgreSQL read-only pool] Idle connection error (connection evicted):', err.message));
             const readOnlyTestClient = await readOnlyPgPool.connect();
             try {
                 for (const warning of await DescribeReadOnlyLoginOverreach(readOnlyTestClient, mj_core_schema)) {
@@ -496,11 +512,19 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                 // long CodeGen timeout rather than the API one, as the SQL Server CodeGen pool does.
                 const codegenPgConnectionConfig = BuildPostgreSQLConnectionConfig({ ...pgEndpoint, User: pgCodegenUser, Password: pgCodegenPass }, configInfo.databaseSettings, 'codegen');
                 const codegenPgPool = new pg.default.Pool(ToPGPoolConfig(codegenPgConnectionConfig));
+                codegenPgPool.on('error', (err) => console.error('[PostgreSQL CodeGen pool] Idle connection error (connection evicted):', err.message));
                 const codegenTestClient = await codegenPgPool.connect();
                 await codegenTestClient.query('SELECT 1');
                 codegenTestClient.release();
                 const { RuntimeSchemaManager } = await import('@memberjunction/schema-engine');
-                const codegenPgConfigData = new PostgreSQLProviderConfigData(codegenPgConnectionConfig, mj_core_schema, cacheRefreshInterval / 1000);
+                // Adopt the already-loaded global Metadata.Provider catalog (ignoreExistingMetadata = false)
+                // so boot does not pay a second full metadata load. The refresh interval stays non-zero on
+                // PostgreSQL on purpose: PostgreSQLDataProvider.AllowRefresh is `CheckRefreshIntervalSeconds > 0`,
+                // so an interval of 0 would make every later codegenPgProvider.Refresh() — the one RSU CodeGen
+                // issues after DDL — return without reloading, leaving this provider frozen at the boot catalog.
+                // (SQL Server's AllowRefresh is independent of the interval, so its codegen provider can use 0.)
+                const codegenPgConfigData = new PostgreSQLProviderConfigData(codegenPgConnectionConfig, mj_core_schema, cacheRefreshInterval / 1000, // ms → seconds
+                undefined, undefined, false);
                 const codegenPgProvider = new PostgreSQLDataProvider();
                 await codegenPgProvider.Config(codegenPgConfigData); // separate pool (per-instance manager); does NOT touch the global API provider
                 RuntimeSchemaManager.Instance.SetDDLProvider(codegenPgProvider);
@@ -523,7 +547,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                     });
                     startupLog.LogIf('verbose', 'RSU in-process CodeGen runner initialized (PostgreSQL).');
                     // Inject CodeGen output paths for targeted git staging
-                    const { initializeConfig } = await import('@memberjunction/codegen-lib');
+                    const { initializeConfig, ResolveAdditionalSchemaInfoPath } = await import('@memberjunction/codegen-lib');
                     const cgConfig = initializeConfig(rsuWorkDir);
                     const outputPaths = (cgConfig.output ?? []).map((o) => o.directory);
                     RuntimeSchemaManager.Instance.SetCodeGenOutputPaths(outputPaths);
@@ -531,9 +555,12 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                     // Point RSU's soft PK/FK writer at the SAME file CodeGen reads (mj.config.cjs
                     // `additionalSchemaInfo`), or RSU writes soft PKs to its own default path while
                     // CodeGen reads a different one and skips integration tables with "No primary key found".
-                    if (cgConfig.additionalSchemaInfo) {
-                        RuntimeSchemaManager.Instance.SetAdditionalSchemaInfoPath(cgConfig.additionalSchemaInfo);
-                        startupLog.LogIf('verbose', `RSU additionalSchemaInfo path: ${cgConfig.additionalSchemaInfo}`);
+                    // Inject the ABSOLUTE path CodeGen resolves (env override, else relative to the mj.config
+                    // file's directory) so RSU, the resolvers and CodeGen agree whatever their cwd (#5858).
+                    const resolvedSchemaInfoPath = ResolveAdditionalSchemaInfoPath();
+                    if (resolvedSchemaInfoPath) {
+                        RuntimeSchemaManager.Instance.SetAdditionalSchemaInfoPath(resolvedSchemaInfoPath);
+                        startupLog.LogIf('verbose', `RSU additionalSchemaInfo path: ${resolvedSchemaInfoPath}`);
                     }
                     else if (RuntimeSchemaManager.Instance.IsEnabled) {
                         // U5 — boot-time assert: RSU is enabled but CodeGen has NO additionalSchemaInfo path
@@ -541,7 +568,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                         // resolved soft PK would be silently lost ("No primary key found" per table). LOUD,
                         // named, and actionable — grep for RSU_ADDITIONAL_SCHEMA_INFO_DIVERGENCE.
                         console.error(`[RSU_ADDITIONAL_SCHEMA_INFO_DIVERGENCE] Runtime Schema Update is ENABLED but mj.config.cjs has no ` +
-                            `'additionalSchemaInfo' setting — RSU writes soft PK/FK definitions to '${process.env.RSU_ADDITIONAL_SCHEMA_INFO_PATH ?? 'additionalSchemaInfo.json'}' ` +
+                            `'additionalSchemaInfo' setting (and RSU_ADDITIONAL_SCHEMA_INFO_PATH is unset) — RSU writes soft PK/FK definitions to '${RuntimeSchemaManager.Instance.AdditionalSchemaInfoPath}' ` +
                             `while CodeGen reads none, so resolved soft primary keys will be silently LOST at codegen time. ` +
                             `Set 'additionalSchemaInfo' in mj.config.cjs to the same file so the write path and read path agree.`);
                     }
@@ -627,8 +654,11 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                 });
                 await codegenPool.connect();
                 const { RuntimeSchemaManager } = await import('@memberjunction/schema-engine');
-                // Same ms→seconds seam as the main provider config above — see providerConfigUnits.ts
-                const codegenConfig = new SQLServerProviderConfigData(codegenPool, mj_core_schema, MetadataCacheRefreshIntervalSeconds(cacheRefreshInterval));
+                // Adopt the already-loaded global Metadata.Provider catalog (ignoreExistingMetadata = false)
+                // and disable a duplicate background refresh timer (0): avoids a second full metadata load
+                // on boot while explicit codegenProvider.Refresh() after DDL still reloads from the DB
+                // (SQLServerDataProvider.AllowRefresh does not depend on the interval; PostgreSQL's does — see above).
+                const codegenConfig = new SQLServerProviderConfigData(codegenPool, mj_core_schema, 0, undefined, undefined, false);
                 const codegenProvider = new SQLServerDataProvider();
                 await codegenProvider.Config(codegenConfig);
                 RuntimeSchemaManager.Instance.SetDDLProvider(codegenProvider);
@@ -651,7 +681,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                     });
                     startupLog.LogIf('verbose', 'RSU in-process CodeGen runner initialized.');
                     // Inject CodeGen output paths for targeted git staging
-                    const { initializeConfig } = await import('@memberjunction/codegen-lib');
+                    const { initializeConfig, ResolveAdditionalSchemaInfoPath } = await import('@memberjunction/codegen-lib');
                     const codegenConfig = initializeConfig(rsuWorkDir);
                     const outputPaths = (codegenConfig.output ?? []).map((o) => o.directory);
                     RuntimeSchemaManager.Instance.SetCodeGenOutputPaths(outputPaths);
@@ -660,14 +690,17 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                     // `additionalSchemaInfo`). Without this, RSU writes soft PKs to its own default
                     // path while CodeGen reads a different one and skips every integration table
                     // with "No primary key found".
-                    if (codegenConfig.additionalSchemaInfo) {
-                        RuntimeSchemaManager.Instance.SetAdditionalSchemaInfoPath(codegenConfig.additionalSchemaInfo);
-                        startupLog.LogIf('verbose', `RSU additionalSchemaInfo path: ${codegenConfig.additionalSchemaInfo}`);
+                    // Inject the ABSOLUTE path CodeGen resolves (env override, else relative to the mj.config
+                    // file's directory) so RSU, the resolvers and CodeGen agree whatever their cwd (#5858).
+                    const resolvedSchemaInfoPath = ResolveAdditionalSchemaInfoPath();
+                    if (resolvedSchemaInfoPath) {
+                        RuntimeSchemaManager.Instance.SetAdditionalSchemaInfoPath(resolvedSchemaInfoPath);
+                        startupLog.LogIf('verbose', `RSU additionalSchemaInfo path: ${resolvedSchemaInfoPath}`);
                     }
                     else if (RuntimeSchemaManager.Instance.IsEnabled) {
                         // U5 — boot-time assert: write path ≠ read path (see the PostgreSQL branch for detail).
                         console.error(`[RSU_ADDITIONAL_SCHEMA_INFO_DIVERGENCE] Runtime Schema Update is ENABLED but mj.config.cjs has no ` +
-                            `'additionalSchemaInfo' setting — RSU writes soft PK/FK definitions to '${process.env.RSU_ADDITIONAL_SCHEMA_INFO_PATH ?? 'additionalSchemaInfo.json'}' ` +
+                            `'additionalSchemaInfo' setting (and RSU_ADDITIONAL_SCHEMA_INFO_PATH is unset) — RSU writes soft PK/FK definitions to '${RuntimeSchemaManager.Instance.AdditionalSchemaInfoPath}' ` +
                             `while CodeGen reads none, so resolved soft primary keys will be silently LOST at codegen time. ` +
                             `Set 'additionalSchemaInfo' in mj.config.cjs to the same file so the write path and read path agree.`);
                     }
@@ -868,6 +901,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // config.ts loads and validates at module scope, so importing it there would pull full config
     // validation into every import chain that touches the resolver, unit tests included.
     ConfigureRecordDataBroadcast(configInfo.cacheSettings?.recordDataBroadcastEntities);
+    CacheInvalidationBroadcastManager.Instance.Configure(configInfo.cacheSettings?.broadcast);
     // Global listener: broadcast CACHE_INVALIDATION to all browser clients whenever
     // ANY BaseEntity save/delete occurs on this server — regardless of whether it
     // originated from a GraphQL mutation or internal server-side code (agents, actions,
@@ -888,7 +922,8 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
                     // Opt-in only: this event reaches every connected client unfiltered, and this listener
                     // fires for server-internal saves too (agents, actions, orchestrator), which are exactly
                     // the ones no browser session asked for.
-                    recordData: beEvent.type === 'save' && MayBroadcastRecordData(entityName)
+                    recordData: beEvent.type === 'save' && MayBroadcastRecordData(entityName, beEvent.baseEntity.EntityInfo)
+                        // fls-guard: MayBroadcastRecordData refuses FLS entities and Record Changes
                         ? JSON.stringify(beEvent.baseEntity.GetAll())
                         : undefined,
                 });
@@ -936,7 +971,6 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
         scalarsMap: [{ type: Date, scalar: GraphQLTimestamp }],
         emitSchemaFile: websiteRunFromPackage !== 1,
         pubSub,
-        globalMiddlewares: [variablesLoggingMiddleware],
     });
     const buildMs = performance.now() - tBuild;
     if (cpuProf && profSession) {
@@ -957,6 +991,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // and aren't @NoLog-marked. No-op in default config (logVariables=false).
     auditResolversForUndecoratedArgs();
     const tTransform = performance.now();
+    schema = ApplyRootVariablesLoggingMiddleware(schema);
     schema = requireSystemUserDirective.transformer(schema);
     schema = publicDirective.transformer(schema);
     // Apply middleware-contributed schema transformers (after built-in directive transformers)
@@ -1066,6 +1101,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
         },
         // context() now just reads the already-validated payload — no I/O, no logging
         context: async (ctx) => {
+            // No RequestOrigin here: the upgrade request's Origin is not carried onto subscription contexts.
             return { userPayload: ctx.extra.userPayload };
         },
         onClose: (ctx) => {
@@ -1124,6 +1160,15 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // ─── Server extensions loader & shared service registry ───────────────────
     const extensionLoader = new ServerExtensionLoader();
     _currentExtensionLoader = extensionLoader;
+    const hostRuntimeService = {
+        CreateIsolatedProvider,
+        GetSystemUser: () => getSystemUser(),
+        ResolvePrincipal: (configuredEmailOrId, subsystemLabel) => ResolveConfiguredPrincipal(configuredEmailOrId, subsystemLabel),
+    };
+    extensionLoader.Services.RegisterService(SERVER_EXTENSION_CORE_SERVICE_KEYS.HostRuntime, hostRuntimeService);
+    if (coreAuthenticatedRouters.MagicLinkService) {
+        extensionLoader.Services.RegisterService(SERVER_EXTENSION_CORE_SERVICE_KEYS.MagicLink, coreAuthenticatedRouters.MagicLinkService);
+    }
     // Open App packages publish their extensions; host mj.config.cjs overlays by DriverClass
     // Backwards-compatibility shim: synthesize ServerExtensionConfig entries from legacy configInfo.telephony
     const telephonyExtensionConfigs = [];
@@ -1226,6 +1271,11 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     MountExtensionHealthRoute(app, extensionLoader);
     MountAuthCatalogRoute(app, startupLog);
     // ─── Unified auth middleware (replaces both REST authMiddleware and contextFunction auth) ─────
+    // Always shown, whatever the log level: a configured map makes logins run as other people's records
+    const userEmailMapWarning = DescribeUserEmailMap(userEmailMap);
+    if (userEmailMapWarning) {
+        console.warn(userEmailMapWarning);
+    }
     app.use(CreateUnifiedAuthMiddleware(dataSources));
     // ─── Post-auth middleware from BaseServerMiddleware plugins ─────
     // Middleware here has access to the authenticated user via req.userPayload.
@@ -1303,11 +1353,29 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // No per-route authMiddleware needed — unified auth middleware already ran
     setupRESTEndpoints(app, restApiConfig);
     // ─── GraphQL middleware (contextFunction reads req.userPayload, no re-auth) ─────
-    app.use(graphqlRootPath, cors(), BodyParser.json({ limit: '50mb' }), 
+    app.use(graphqlRootPath, cors({
+        exposedHeaders: ['X-MJ-Cache-Seq', 'X-MJ-Server-ID'],
+    }), BodyParser.json({ limit: '50mb' }), 
     // Express 5 leaves req.body as undefined for non-JSON or empty bodies;
     // Apollo Server's expressMiddleware requires req.body to be defined.
-    (req, _res, next) => { if (req.body === undefined)
-        req.body = {}; next(); }, expressMiddleware(apolloServer, {
+    // Also piggyback the latest monotonic cache invalidation sequence and server process ID
+    // onto every GraphQL HTTP response so clients can detect and recover missed invalidations.
+    (req, res, next) => {
+        if (req.body === undefined)
+            req.body = {};
+        const broadcastProvider = CacheInvalidationBroadcastManager.Instance.Provider;
+        res.setHeader('X-MJ-Cache-Seq', String(broadcastProvider.CurrentSequence));
+        res.setHeader('X-MJ-Server-ID', broadcastProvider.ServerProcessID);
+        const originalWriteHead = res.writeHead.bind(res);
+        res.writeHead = ((...args) => {
+            if (!res.headersSent) {
+                res.setHeader('X-MJ-Cache-Seq', String(broadcastProvider.CurrentSequence));
+                res.setHeader('X-MJ-Server-ID', broadcastProvider.ServerProcessID);
+            }
+            return originalWriteHead(...args);
+        });
+        next();
+    }, expressMiddleware(apolloServer, {
         context: contextFunction({
             setupComplete$,
             dataSource: extendConnectionPoolWithQuery(dataSources[0].dataSource), // default read-write data source
@@ -1318,9 +1386,34 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     for (const mw of mwPostRoute) {
         app.use(mw);
     }
+    const serverRole = ResolveServerRole(configInfo.serverRole ?? process.env.MJ_SERVER_ROLE);
+    const runBackgroundWorkers = ShouldRunBackgroundWorkers(serverRole);
+    if (!runBackgroundWorkers) {
+        LogStatus(`[ServerRole] Running in '${serverRole}' mode — background worker services are disabled on this replica.`);
+    }
+    // Every role runs agents, so every role's watchdog takes the configured intervals (before any run starts).
+    AgentRunWatchdog.Instance.Configure(configInfo.agentRunWatchdog);
+    // The handoff offer registry starts its own expiry sweep when it loads, on every role; apply the host's setting.
+    ApplyHandoffOfferExpirySettings(configInfo.agentSessions.handoffOfferExpiry);
+    // Close out run artifacts a dead process left open, BEFORE anything below can start a new run.
+    //
+    // A killed process (OOM killer, SIGKILL, host reboot) never writes its run's terminal result, so
+    // the run reads as in-flight forever: the wizard spins, the active-operations panel shows a step
+    // that will never advance, and no failure is reported anywhere. Sound only before the first new
+    // run: artifacts are process-local, so anything still in flight as THIS process starts belongs
+    // to a process that no longer exists. That is why it runs here, ahead of the scheduler, the
+    // integration sync worker (its Start() polls immediately), the HTTP listener (a request can
+    // start a discovery) and pending RSU work, and why it is awaited. It needs none of them.
+    //
+    // One exception: an RSU run killed inside its own RestartMJAPI step is not dead — THIS process is
+    // its continuation, and processRSUPendingWork below re-attaches to it. Those are left in flight
+    // (and named in `Skipped`); whichever of them nothing re-attaches to is closed out afterwards by
+    // RSUPostRestartProgressSession.SweepUnattached. On a replica that runs no background workers
+    // nothing would ever resume them, so there they are swept like any other.
+    const bootSweep = await SweepInterruptedRuns(undefined, msg => console.warn(msg), runBackgroundWorkers ? RSUPostRestartProgressSession.WasKilledAtRestart : undefined);
     // Initialize and start scheduled jobs service if enabled
     let scheduledJobsService = null;
-    if (configInfo.scheduledJobs?.enabled) {
+    if (runBackgroundWorkers && configInfo.scheduledJobs?.enabled) {
         try {
             scheduledJobsService = new ScheduledJobsService(configInfo.scheduledJobs);
             await scheduledJobsService.Initialize();
@@ -1334,7 +1427,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // Initialize and start the integration sync worker if enabled (PR 1 item 8).
     // Off by default — deployments that don't opt in keep running syncs inline.
     let integrationSyncWorker = null;
-    if (configInfo.integrationSyncWorker?.enabled) {
+    if (runBackgroundWorkers && configInfo.integrationSyncWorker?.enabled) {
         try {
             integrationSyncWorker = new IntegrationSyncWorkerService(configInfo.integrationSyncWorker);
             await integrationSyncWorker.Initialize();
@@ -1368,47 +1461,44 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
         startupLog.SetScheduledJobCount(scheduledJobsService.GetStatus().activeJobs);
     }
     startupLog.PrintSummary();
-    // Process pending RSU work from pre-restart (entity maps, field maps, sync)
-    processRSUPendingWork().catch(err => console.warn(`RSU pending work processing failed: ${err}`));
-    // Resume any integration syncs that were orphaned by the previous process restart
     const resumeUser = UserCache.Instance.GetSystemUser();
-    if (resumeUser) {
-        IntegrationEngine.Instance.ResumeOrphanedSyncs(resumeUser)
-            .catch(err => console.warn(`[IntegrationEngine] Orphaned sync resume failed: ${err}`));
-    }
-    // Force-fail any agent runs left 'Running' by a process that died (restart/crash/OOM) or whose
-    // terminal-state write never landed. Staleness-based, so it never touches runs another healthy
-    // instance is still heart-beating. The watchdog also self-registers for graceful-shutdown
-    // cancellation (via ShutdownRegistry) once it begins tracking this process's first live run.
-    if (resumeUser && Metadata.Provider instanceof DatabaseProviderBase) { // global-provider-ok: server startup recovery — one-shot orphaned-run sweep at boot
-        const sweepUser = resumeUser;
-        const sweepProvider = Metadata.Provider; // global-provider-ok: server startup recovery — one-shot orphaned-run sweep at boot
-        AgentRunWatchdog.SweepOrphanedRuns(sweepProvider, sweepUser)
-            .catch(err => console.warn(`[AgentRunWatchdog] Startup sweep failed: ${err}`));
-        // The watchdog repairs the RUN; nothing repaired the conversation detail, which is the row the
-        // chat actually renders from (MJ #4222). A process that dies mid-run leaves a terminal run
-        // beside a message that still claims to be generating, and it spins forever for anyone who
-        // opens it. Runs at boot (closes restart orphans) and on a timer (closes mid-life orphans),
-        // mirroring the watchdog's own two-phase shape.
-        const reconcileOrphans = () => ReconcileOrphanedConversationDetails(sweepProvider, sweepUser)
-            .catch(err => console.warn(`[OrphanDetailReconciler] Pass failed: ${err}`));
-        void reconcileOrphans();
-        const orphanDetailTimer = setInterval(() => void reconcileOrphans(), ORPHAN_DETAIL_SWEEP_INTERVAL_MS);
-        ShutdownRegistry.Instance.Register({
-            ShutdownName: 'OrphanDetailReconciler',
-            Shutdown: () => { clearInterval(orphanDetailTimer); },
-        });
-    }
-    // Launch the AI Agent Session janitor: run own-host orphan recovery once at boot, then a periodic tick
-    // that closes stale and over-deadline sessions and rebuilds lost realtime recordings from their
-    // crash-recovery shards (storage reads and writes). Self-registers with ShutdownRegistry, so its timer
-    // is cleared by the gracefulShutdown drain below (no explicit Stop() wiring needed here).
-    if (resumeUser && Metadata.Provider instanceof DatabaseProviderBase) { // global-provider-ok: server startup recovery — boot-time session janitor uses the server's own provider
-        SessionJanitor.Instance.Start(Metadata.Provider, resumeUser) // global-provider-ok: server-owned background reconciler runs under the server's provider + system user
-            .catch(err => console.warn(`[SessionJanitor] Startup failed: ${err}`));
-        // The bridge engine's orphan reconcile likewise starts only now that this process holds its port, so it can't
-        // take the live bridges of another MJAPI still serving that port for its own orphans (#5310).
-        StartBridgeOrphanReconciliation(Metadata.Provider, resumeUser); // global-provider-ok: server-owned background reconciler runs under the server's provider + system user
+    if (runBackgroundWorkers) {
+        // Process pending RSU work from pre-restart (entity maps, field maps, sync)
+        processRSUPendingWork().catch(err => console.warn(`RSU pending work processing failed: ${err}`))
+            .finally(() => RSUPostRestartProgressSession.SweepUnattached(bootSweep.Skipped, undefined, msg => console.warn(msg)).catch(err => console.warn(`RSU leftover run sweep failed: ${err}`)));
+        // Resume any integration syncs that were orphaned by the previous process restart
+        if (resumeUser) {
+            IntegrationEngine.Instance.ResumeOrphanedSyncs(resumeUser)
+                .catch(err => console.warn(`[IntegrationEngine] Orphaned sync resume failed: ${err}`));
+        }
+        // Force-fail any agent runs left 'Running' by a process that died (restart/crash/OOM) or whose
+        // terminal-state write never landed. Staleness-based, so it never touches runs another healthy
+        // instance is still heart-beating. The watchdog also self-registers for graceful-shutdown
+        // cancellation (via ShutdownRegistry) once it begins tracking this process's first live run.
+        if (resumeUser && Metadata.Provider instanceof DatabaseProviderBase) { // global-provider-ok: server startup recovery — one-shot orphaned-run sweep at boot
+            const sweepUser = resumeUser;
+            const sweepProvider = Metadata.Provider; // global-provider-ok: server startup recovery — one-shot orphaned-run sweep at boot
+            AgentRunWatchdog.SweepOrphanedRuns(sweepProvider, sweepUser, configInfo.agentRunWatchdog)
+                .catch(err => console.warn(`[AgentRunWatchdog] Startup sweep failed: ${err}`));
+            // The watchdog repairs the RUN; nothing repaired the conversation detail, which is the row the
+            // chat actually renders from (MJ #4222). A process that dies mid-run leaves a terminal run
+            // beside a message that still claims to be generating, and it spins forever for anyone who
+            // opens it. Runs at boot (closes restart orphans) and, unless `conversationDetailReconciler.enabled`
+            // is false, on a timer (closes mid-life orphans), mirroring the watchdog's own two-phase shape.
+            StartOrphanedConversationDetailReconciler(sweepProvider, sweepUser, configInfo.conversationDetailReconciler);
+        }
+        // Launch the AI Agent Session janitor: run own-host orphan recovery once at boot, then a periodic tick
+        // that closes stale and over-deadline sessions and rebuilds lost realtime recordings from their
+        // crash-recovery shards (storage reads and writes). Self-registers with ShutdownRegistry, so its timer
+        // is cleared by the gracefulShutdown drain below (no explicit Stop() wiring needed here).
+        if (resumeUser && Metadata.Provider instanceof DatabaseProviderBase) { // global-provider-ok: server startup recovery — boot-time session janitor uses the server's own provider
+            // `agentSessions.janitor.enabled: false` keeps the boot recovery and shutdown drain but schedules no tick.
+            SessionJanitor.Instance.StartWithSettings(configInfo.agentSessions.janitor, Metadata.Provider, resumeUser) // global-provider-ok: server-owned background reconciler runs under the server's provider + system user
+                .catch(err => console.warn(`[SessionJanitor] Startup failed: ${err}`));
+            // The bridge engine's orphan reconcile likewise starts only now that this process holds its port, so it can't
+            // take the live bridges of another MJAPI still serving that port for its own orphans (#5310).
+            StartBridgeOrphanReconciliation(Metadata.Provider, resumeUser, configInfo.agentSessions.bridgeReconciler); // global-provider-ok: server-owned background reconciler runs under the server's provider + system user
+        }
     }
     // Launch the durable task-graph dispatcher: claim Pending tasks, execute them with a fresh
     // provider each, and reconcile claims orphaned by a crash. Without this nothing ever picks up a
@@ -1437,7 +1527,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
         SuppressTaskGraphSubmission('MJ_DISABLE_TASK_GRAPH_DISPATCHER=1 is set on this host');
         LogStatus('[TaskGraphDispatcher] Disabled by MJ_DISABLE_TASK_GRAPH_DISPATCHER=1 — this process will neither accept nor execute task graphs.');
     }
-    else if (resumeUser && taskGraphPool instanceof sql.ConnectionPool) {
+    else if (runBackgroundWorkers && resumeUser && taskGraphPool instanceof sql.ConnectionPool) {
         StartTaskGraphDispatcher(taskGraphPool, resumeUser)
             .catch(err => console.warn(`[TaskGraphDispatcher] Startup failed: ${err}`));
     }
@@ -1446,7 +1536,7 @@ export const Serve = async (resolverPaths, app = CreateApp(), options) => {
     // ShutdownAll() drains it (up to 2 × shutdownDrainMs) before the HTTP server closes.
     // Not awaited: a slow engine load must not delay readiness, and a failure never stops the API.
     const workQueueProvider = Metadata.Provider; // global-provider-ok: server startup — the work-queue host runs on the server's own provider
-    if (configInfo.workQueue?.enabled && workQueueProvider instanceof DatabaseProviderBase) {
+    if (runBackgroundWorkers && configInfo.workQueue?.enabled && workQueueProvider instanceof DatabaseProviderBase) {
         const workQueuePool = dataSources[0]?.dataSource;
         const providerSource = new MJServerWorkQueueProviderSource(workQueuePool instanceof sql.ConnectionPool ? workQueuePool : null, workQueueProvider);
         StartWorkQueueHost(configInfo.workQueue, workQueueProvider, providerSource)
@@ -1589,9 +1679,17 @@ async function processRSUPendingWork() {
     console.log(`[RSU] Processing ${pendingItems.length} pending work item(s) from pre-restart...`);
     // Wait a moment for metadata to be fully loaded
     await new Promise(resolve => setTimeout(resolve, 3000));
+    // The RSU run these items belong to is still open: the pre-restart process was killed inside its
+    // own RestartMJAPI step, so it never wrote a terminal result. This session re-attaches to that run
+    // and reports the two steps that only ever happen here — CreateEntityMaps and StartSync — into the
+    // SAME event stream, so a client that has been tailing since before the restart keeps its cursor
+    // and finally learns whether the connector went live. Purely observational: every failure below is
+    // swallowed, and the work proceeds whether or not a run was found to report into.
+    const progressSession = new RSUPostRestartProgressSession();
     for (const pending of pendingItems) {
         const pendingWorkID = pending.ID;
         const item = pending.Work;
+        let progress = null;
         // Declared outside the try so the catch can narrow a retry to what is still outstanding: what
         // actually got mapped this attempt. Each retry is then strictly smaller, and one poison object
         // cannot keep re-running its healthy siblings.
@@ -1610,6 +1708,9 @@ async function processRSUPendingWork() {
                 await ProcessPromoteColumnsPendingWork(item, pendingWorkID, rsm, systemUser);
                 continue;
             }
+            // Attached only for apply-objects work: promote-columns is a follow-up to a DIFFERENT run
+            // shape that has no CreateEntityMaps/StartSync steps to report.
+            progress = await progressSession.For(item.CompanyIntegrationID);
             // Resolve connector
             const rv = new RunView();
             const ciResult = await rv.RunView({
@@ -1619,8 +1720,10 @@ async function processRSUPendingWork() {
             }, systemUser);
             const companyIntegration = ciResult.Results[0];
             if (!companyIntegration) {
-                await rsm.FailPendingWork(pendingWorkID, `CompanyIntegration ${item.CompanyIntegrationID} not found`, systemUser);
-                console.warn(`[RSU] CompanyIntegration ${item.CompanyIntegrationID} not found`);
+                const missingCI = `CompanyIntegration ${item.CompanyIntegrationID} not found`;
+                progress?.FailEntityMaps(missingCI, item.SourceObjectNames ?? []);
+                await rsm.FailPendingWork(pendingWorkID, missingCI, systemUser);
+                console.warn(`[RSU] ${missingCI}`);
                 continue;
             }
             const integrationName = companyIntegration.Integration;
@@ -1631,23 +1734,39 @@ async function processRSUPendingWork() {
             }, systemUser);
             const integrationEntity = integrationResult.Results[0];
             if (!integrationEntity) {
-                await rsm.FailPendingWork(pendingWorkID, `Integration entity for ${integrationName} not found`, systemUser);
-                console.warn(`[RSU] Integration entity for ${integrationName} not found`);
+                const missingIntegration = `Integration entity for ${integrationName} not found`;
+                progress?.FailEntityMaps(missingIntegration, item.SourceObjectNames ?? []);
+                await rsm.FailPendingWork(pendingWorkID, missingIntegration, systemUser);
+                console.warn(`[RSU] ${missingIntegration}`);
                 continue;
             }
             const connector = ConnectorFactory.Resolve(integrationEntity);
             if (!connector) {
-                await rsm.FailPendingWork(pendingWorkID, `Connector for ${integrationName} not found`, systemUser);
-                console.warn(`[RSU] Connector for ${integrationName} not found`);
+                const missingConnector = `Connector for ${integrationName} not found`;
+                progress?.FailEntityMaps(missingConnector, item.SourceObjectNames ?? []);
+                await rsm.FailPendingWork(pendingWorkID, missingConnector, systemUser);
+                console.warn(`[RSU] ${missingConnector}`);
                 continue;
             }
             // Create entity maps + field maps for each source object
             const createdEntityMapIDs = [];
             const rvPending = new RunView();
             const sourceObjectFields = item.SourceObjectFields ?? {};
-            // Introspect schema ONCE for the entire connector, then reuse per object
+            // Introspect schema ONCE for the entire connector, then reuse per object.
+            //
+            // IN THE CONNECTION'S CATALOG SCOPE. This runs at process start, outside any request, so
+            // nothing upstream has entered a scope for it — and an unscoped catalog read falls back to
+            // the SHARED catalog by design. That would build these field maps from the connector's
+            // declared fields alone: every column this connection's own discovery sampled would get a
+            // table column from the RSU (built in scope, from the per-connection rows) but no field map,
+            // and its values would land in the overflow JSON instead of the column. Observed on the
+            // sandbox 2026-09-11: 487 declared fields, 556 per-connection — 69 columns that this read
+            // would have silently left unmapped. Entering the scope is not enough on a fresh process:
+            // IntrospectInCatalogScope loads the catalog and warms each object's fields first.
             const introspect = connector.IntrospectSchema.bind(connector);
-            const schema = await introspect(companyIntegration, systemUser);
+            const schema = await IntrospectInCatalogScope(item.CompanyIntegrationID, item.SourceObjectNames, () => introspect(companyIntegration, systemUser), systemUser);
+            const objectsToMap = item.SourceObjectNames.length;
+            progress?.BeginEntityMaps(objectsToMap);
             for (const objName of item.SourceObjectNames) {
                 const tableName = objName.replace(/[^A-Za-z0-9_]/g, '_').toLowerCase();
                 const entity = md.Entities.find(e => e.SchemaName.toLowerCase() === item.SchemaName.toLowerCase() &&
@@ -1742,7 +1861,12 @@ async function processRSUPendingWork() {
                 catch (fieldErr) {
                     console.warn(`[RSU] Field map creation failed for ${objName}: ${fieldErr}`);
                 }
+                progress?.ObjectMapped(objName, mappedObjectNames.size, objectsToMap);
             }
+            // The applied rollup for this half of the run, in OBJECTS. `mappedObjectNames` is the same set
+            // the retry path narrows against, so what is reported succeeded here is exactly what a retry
+            // would NOT redo.
+            progress?.CompleteEntityMaps(objectsToMap, mappedObjectNames.size, objectsToMap - mappedObjectNames.size);
             // Remove-as-disable: entity maps whose object is NOT in this apply's selection
             // are disabled (data kept; re-selection re-enables). 'ignore' opts out for subset applies.
             if ((item.UnselectedAction ?? 'disable') !== 'ignore') {
@@ -1757,8 +1881,11 @@ async function processRSUPendingWork() {
                     console.warn(`[RSU] Unselected-map disable failed: ${disableErr}`);
                 }
             }
-            // Start sync if requested
+            // Start sync if requested. This is the step the setup journey is actually waiting on — the
+            // connector is not live until it begins — so it is a stage of its own rather than a footnote
+            // on entity-map creation.
             if (item.StartSync !== false) {
+                progress?.BeginStartSync();
                 try {
                     await IntegrationEngine.Instance.Config(false, systemUser);
                     const syncOptions = {};
@@ -1771,13 +1898,20 @@ async function processRSUPendingWork() {
                     const opts = Object.keys(syncOptions).length > 0 ? syncOptions : undefined;
                     IntegrationEngine.Instance.RunSync(item.CompanyIntegrationID, systemUser, 'Manual', undefined, undefined, opts);
                     console.log(`[RSU] Sync started for ${item.CompanyIntegrationID} (EntityMaps: ${createdEntityMapIDs.length}, FullSync: ${!!item.FullSync}, SyncDirection: ${item.SyncDirection ?? 'entity-map default'})`);
+                    // Looked up rather than returned: RunSync resolves only when the whole sync finishes, and
+                    // this stage closes when the sync STARTS. A null id closes the stage anyway — the sync did
+                    // start; only the hand-off to its own stream is missing.
+                    progress?.CompleteStartSync(await FindStartedSyncRunID(item.CompanyIntegrationID, systemUser));
                 }
                 catch (syncErr) {
                     console.warn(`[RSU] Sync start failed: ${syncErr}`);
+                    progress?.FailStartSync(`Sync start failed: ${syncErr instanceof Error ? syncErr.message : String(syncErr)}`);
                 }
             }
             else {
                 console.log(`[RSU] Sync skipped for ${item.CompanyIntegrationID} (StartSync=false)`);
+                progress?.BeginStartSync();
+                progress?.SkipStartSync('the caller asked for no initial sync (StartSync=false)');
             }
             // Create or update schedule if CronExpression provided
             if (item.CronExpression) {
@@ -1813,6 +1947,7 @@ async function processRSUPendingWork() {
                             EntityName: 'MJ: Scheduled Job Types',
                             ExtraFilter: `DriverClass='IntegrationSyncScheduledJobDriver'`,
                             MaxRows: 1,
+                            SkipTotalRowCount: true,
                             ResultType: 'simple',
                             Fields: ['ID']
                         }, systemUser);
@@ -1861,6 +1996,16 @@ async function processRSUPendingWork() {
             // Failing terminally on the first error means the objects this item would have mapped are
             // silently never mapped, and the only recovery is someone noticing and re-applying by hand.
             const remaining = (item.SourceObjectNames ?? []).filter(n => !mappedObjectNames.has(n));
+            // Same list the retry narrows to, reported into the run stream: when this half fails, those
+            // objects have tables and no way to sync into them, and this is the client's only signal.
+            //
+            // Reported under CreateEntityMaps because that is where the overwhelming majority of this
+            // block's work — and so its failures — lives. A throw from further down (schedule creation,
+            // say) lands here too, with an empty `remaining`; the stage attribution is then approximate
+            // but the message is the real error and the run is correctly marked failed, which is what the
+            // operator needs. The alternative — inventing a stage that is not in the pipeline's step list
+            // — would break the determinate stepper for the sake of a nicer label.
+            progress?.FailEntityMaps(message, remaining);
             const requeued = await rsm.RetryPendingWork(pendingWorkID, item, remaining, systemUser);
             if (!requeued) {
                 // Budget spent, or nothing left to retry. This message is the operator's only signal, so
@@ -1870,8 +2015,40 @@ async function processRSUPendingWork() {
             }
         }
     }
+    // Terminates every RSU run this pass re-attached to. Until this lands, those runs have been
+    // flagged in-flight since before the restart — which is exactly the state this whole mechanism
+    // exists to end.
+    await progressSession.FinishAll();
     console.log(`[RSU] Pending work processing complete`);
 }
+/**
+ * The ID of the sync run a just-launched sync created, or null if it is not readable yet.
+ *
+ * `RunSync` resolves only when the entire sync FINISHES, so it cannot be awaited for this — the
+ * StartSync stage closes when the sync begins, not when it ends. The run row is written early in
+ * the sync's own startup, so a short settle plus a newest-first lookup finds it. Null is a normal
+ * outcome under load and is reported as such; it never fails anything.
+ */
+async function FindStartedSyncRunID(companyIntegrationID, systemUser) {
+    try {
+        await new Promise(resolve => setTimeout(resolve, 200));
+        const rv = new RunView();
+        const runResult = await rv.RunView({
+            EntityName: 'MJ: Company Integration Runs',
+            ExtraFilter: `CompanyIntegrationID='${companyIntegrationID.replace(/'/g, "''")}' AND Status='In Progress'`,
+            OrderBy: '__mj_CreatedAt DESC',
+            MaxRows: 1,
+            SkipTotalRowCount: true,
+            ResultType: 'simple',
+            Fields: ['ID'],
+        }, systemUser);
+        return runResult.Success && runResult.Results.length > 0 ? runResult.Results[0].ID : null;
+    }
+    catch (e) {
+        console.warn(`[RSU] Could not resolve the started sync run id for ${companyIntegrationID}: ${e}`);
+        return null;
+    }
+}
 /**
  * Creates a MSSQL ConnectionPool-compatible wrapper around a pg.Pool.
  * This allows existing code that references DataSourceInfo.dataSource (typed as sql.ConnectionPool)
@@ -1911,10 +2088,11 @@ function mountRealtimeRecordingRoutes(app) {
     }
     const graceMs = () => configInfo.realtime.recordingAssemblyGraceSeconds * 1000;
     const systemUser = UserCache.Instance.GetSystemUser();
-    if (systemUser) {
+    const runWorkers = ShouldRunBackgroundWorkers(ResolveServerRole(configInfo.serverRole ?? process.env.MJ_SERVER_ROLE));
+    if (systemUser && runWorkers) {
         RecordingAssemblyScheduler.Instance.Start({ Provider: provider, SystemUser: systemUser, GraceMs: graceMs });
     }
-    else {
+    else if (!systemUser) {
         LogError('[Realtime] Recording assembly scheduler not started: no system user; the session janitor assembles recordings instead');
     }
     app.use(ApiRoute(REALTIME_RECORDING_MOUNT_PATH), cors(), CreateRealtimeRecordingRouter({
@@ -1923,7 +2101,7 @@ function mountRealtimeRecordingRoutes(app) {
             return user ? { ContextUser: user, Provider: provider } : null;
         },
         LoadOwnedSession: LoadOwnedRealtimeSession,
-        LoadConversationExternalID: LoadSessionConversationExternalID,
+        AuthorizeSession: AuthorizeLoadedSession,
         CreateLedger: (ledgerProvider, contextUser) => new SqlRealtimeRecordingLedger(ledgerProvider, contextUser),
         Scheduler: RecordingAssemblyScheduler.Instance,
         GraceMs: graceMs,