@spooky-sync/core 0.0.1-canary.20 → 0.0.1-canary.201

package/src/modules/sync/sync.ts

@@ -1,33 +1,203 @@
-import { LocalDatabaseService, RemoteDatabaseService } from '../../services/database/index';
-import { MutationEvent, RecordVersionArray } from '../../types';
+import type {
+  ConnectionSupervisor,
+  LocalStore,
+  RemoteDatabaseService,
+} from '../../services/database/index';
+import type {
+  ConnectionState,
+  RecordVersionArray,
+  RecordVersionDiff,
+  SyncHealth,
+  SyncHealthStatus,
+} from '../../types';
 import { createSyncEventSystem, SyncEventTypes, SyncQueueEventTypes } from './events/index';
-import { Logger } from '../../services/logger/index';
-import { DownEvent, DownQueue, UpEvent, UpQueue } from './queue/index';
-import { RecordId, Uuid } from 'surrealdb';
-import { ArraySyncer, createDiffFromDbOp } from './utils';
+import type { Logger } from '../../services/logger/index';
+import type { DownEvent, UpEvent } from './queue/index';
+import { DownQueue, UpQueue } from './queue/index';
+import type { RecordId, Uuid } from 'surrealdb';
+import {
+  applyRecordVersionDiff,
+  ArraySyncer,
+  buildListRefSelect,
+  buildQueryRowCountSelect,
+  buildSubqueryListRefSelect,
+  createDiffFromDbOp,
+  diffRecordVersionArray,
+  listRefPollDelayMs,
+  recordVersionArraysEqual,
+  resolveListRefPollInterval,
+} from './utils';
 import { SyncEngine } from './engine';
 import { SyncScheduler } from './scheduler';
-import { SchemaStructure } from '@spooky-sync/query-builder';
-import { CacheModule } from '../cache/index';
-import { DataModule } from '../data/index';
-import { encodeRecordId, extractTablePart, parseDuration, surql } from '../../utils/index';
+import type { SchemaStructure } from '@spooky-sync/query-builder';
+import type { CacheModule } from '../cache/index';
+import type { DataModule } from '../data/index';
+import {
+  classifySyncError,
+  encodeRecordId,
+  extractIdPart,
+  extractTablePart,
+  surql,
+  withTimeout,
+} from '../../utils/index';
+import { ANON_USER_ID, DEFAULT_REF_MODE, listRefTableFor, RefMode } from '../ref-tables';
+import { mutationOwnerTabId } from '../data/mutation-id';
+import type { LeaderSyncHub, SyncForwarder } from '../../services/tabs/coordinator';
+import { parseRecordIdString } from '../../utils/index';
 
 /**
- * The main synchronization engine for Spooky.
+ * Tunables for `Sp00kySync` construction.
+ */
+export interface Sp00kySyncOptions {
+  /**
+   * Cadence (ms) for the `_00_list_ref` poll fallback that catches
+   * cross-session UPDATEs the LIVE-permission gap drops. Non-positive
+   * values fall back to the default; see
+   * {@link resolveListRefPollInterval}.
+   */
+  refSyncIntervalMs?: number;
+  /**
+   * Enable realtime sync for unauthenticated clients against the shared
+   * `_00_list_ref_anon` table. See {@link Sp00kyConfig.enableAnonymousLiveQueries}.
+   * Defaults to `false`.
+   */
+  anonymousLiveQueries?: boolean;
+  /**
+   * Consecutive failed sync rounds before sync health flips to `degraded`.
+   * `0` disables degraded reporting. See {@link Sp00kyConfig.syncHealth}.
+   * Defaults to `3`.
+   */
+  degradeAfterConsecutiveFailures?: number;
+  /**
+   * Max time a single mutation push may take before it is treated as a network
+   * failure and retried. Guards against an RPC that never settles wedging the
+   * up-queue for the session. Defaults to 30000; `0` disables the timeout.
+   */
+  pushTimeoutMs?: number;
+  /**
+   * Transport supervisor. Sync reads its state to report `connection` in
+   * {@link SyncHealth} so a UI can show "reconnecting…" the instant the socket
+   * drops, without waiting for the degrade threshold. Optional: omitted in
+   * tests, where `connection` then reports `connected`.
+   */
+  connectionSupervisor?: ConnectionSupervisor;
+}
+
+/**
+ * The main synchronization engine for Sp00ky.
  * Handles the bidirectional synchronization between the local database and the remote backend.
  * Uses a queue-based architecture with 'up' (local to remote) and 'down' (remote to local) queues.
  * @template S The schema structure type.
  */
-export class SpookySync<S extends SchemaStructure> {
-  private clientId: string = '';
+export class Sp00kySync<S extends SchemaStructure> {
   private upQueue: UpQueue;
   private downQueue: DownQueue;
   private isInit: boolean = false;
   private logger: Logger;
   private syncEngine: SyncEngine;
+  /** Engine-level events (e.g. `SYNC_REMOTE_DATA_INGESTED`). Distinct
+   *  from `this.events`, which carries Sp00kySync-level events like
+   *  `SYNC_QUERY_UPDATED` and `SYNC_MUTATION_ROLLED_BACK`. */
+  public get engineEvents() {
+    return this.syncEngine.events;
+  }
   private scheduler: SyncScheduler;
+  /**
+   * Set by any event that means the socket we registered on is gone, so the
+   * next `connected` knows it must re-subscribe rather than treat itself as the
+   * initial connect. See {@link subscribeToReconnect}.
+   */
+  private needsResubscribe: boolean = false;
+  /** When the last reconnect-driven full refetch ran, for burst coalescing. */
+  private lastReconnectRefetchAt = 0;
+  /**
+   * Minimum gap between reconnect-driven full refetches. Long enough to absorb
+   * a flapping socket (the SDK reconnect ladder starts at 1s), short enough
+   * that a genuine drop minutes later still refetches.
+   */
+  private static readonly RECONNECT_REFETCH_COOLDOWN_MS = 10_000;
   public events = createSyncEventSystem();
 
+  // Auth identity that drives per-user `_00_list_ref_user_<id>` routing
+  // in `RefMode.Dedicated`. Updated by `setCurrentUserId` from the auth
+  // subscription in `Sp00kyClient`; null when unauthenticated.
+  private currentUserId: string | null = null;
+
+  // ---- shared-tabs role state ----
+  // Followers keep their own remote WS (registration, per-query sync, poll)
+  // but must never drain the shared outbox or hold a second list_ref LIVE.
+  // The leader relays its LIVE events and routes rollbacks by mutation owner.
+  private tabRole: 'solo' | 'leader' | 'follower' = 'solo';
+  private tabId: string | null = null;
+  private hub: LeaderSyncHub | null = null;
+  private forwarder: SyncForwarder | null = null;
+
+  private refMode: RefMode = DEFAULT_REF_MODE;
+
+  // When true, an unauthenticated client still runs the `_00_list_ref` poll
+  // and LIVE subscription, routed to the shared `_00_list_ref_anon` table, so
+  // a logged-out page gets realtime `useQuery` updates. Off by default.
+  private readonly anonLiveEnabled: boolean;
+
+  // Bookkeeping for the LIVE subscription on `_00_list_ref[_user_*]`.
+  // SurrealDB binds the permission context at LIVE-registration time and
+  // the table name in dedicated mode depends on the authenticated user,
+  // so we have to re-register whenever auth state flips.
+  private currentLiveQueryUuid: Uuid | null = null;
+  private liveQueryUnsubscribe: (() => void) | null = null;
+
+  // Periodic re-poll of `_00_list_ref` as a safety net for missed LIVE
+  // notifications. SurrealDB v3 occasionally drops LIVE deliveries
+  // across sessions even when the row matches the permission rule;
+  // this catches those without requiring users to reload. The
+  // interval is configurable via the constructor; see
+  // `resolveListRefPollInterval` for fallback semantics.
+  //
+  // Self-rescheduling rather than setInterval so each tick can pick
+  // its own delay via `nextPollDelayMs` — slows the poll down when
+  // LIVE is delivering events and speeds it back up when LIVE quiets.
+  private listRefPollTimer: ReturnType<typeof setTimeout> | null = null;
+  private listRefPollRunning: boolean = false;
+  // The currently-executing poll tick, if any. `stopListRefPoll` only stops
+  // future ticks; a bucket switch must also AWAIT the in-flight one so its
+  // local writes land in the store it started against.
+  private listRefPollInFlight: Promise<void> | null = null;
+  public readonly refSyncIntervalMs: number;
+
+  // Consecutive poll cycles that observed NO list_ref change. Drives the
+  // adaptive backoff in `startListRefPoll` via `listRefPollDelayMs`: an idle
+  // page coasts from the fast base cadence toward the 5s cap, and any activity
+  // (a poll-detected change or a LIVE event) resets it to 0 so the poll snaps
+  // back to responsive. Replaces the old LIVE-liveness backoff, which kept the
+  // poll pinned at 500ms forever whenever LIVE wasn't firing (the common case
+  // on a quiet page, thanks to the cross-session LIVE-permission gap).
+  private listRefIdleStreak: number = 0;
+
+  // `${queryHash}:${recordId}` -> consecutive rounds the id has been "still
+  // remote" (left the query's list_ref but still exists upstream). Used to
+  // distinguish a PERSISTENT view-membership disagreement (the `job:` churn,
+  // converged once it crosses the threshold) from a record that's merely
+  // mid-deletion (still-remote for ~one round, then gone) — which must NOT be
+  // converged, or it gets stranded in this window before its delete is observed.
+  private stillRemoteStreaks: Map<string, number> = new Map();
+
+  // Wall-clock timestamp (ms) of the most recent LIVE event delivered
+  // through `handleRemoteListRefChange`. Kept as a diagnostic / liveness
+  // signal; the poll cadence is now driven by `listRefIdleStreak`.
+  private lastLiveEventAt: number | null = null;
+
+  // Number of times the initial `_00_list_ref[_user_*]` LIVE subscription
+  // had to retry on `setCurrentUserId`. Stays at 0 when the SSP has
+  // pre-emptively created the user's dedicated tables; otherwise
+  // increments on each retry attempt until LIVE succeeds or attempts
+  // are exhausted. Surfaced as a diagnostic so the e2e suite can prove
+  // the pre-emptive table-creation path is keeping the first sign-in
+  // off the lazy-creation race.
+  private _liveRetryCount: number = 0;
+  public get liveRetryCount(): number {
+    return this._liveRetryCount;
+  }
+
   get isSyncing() {
     return this.scheduler.isSyncing;
   }
@@ -37,13 +207,11 @@ export class SpookySync<S extends SchemaStructure> {
   }
 
   subscribeToPendingMutations(cb: (count: number) => void): () => void {
-    const id1 = this.upQueue.events.subscribe(
-      SyncQueueEventTypes.MutationEnqueued,
-      (event) => cb(event.payload.queueSize)
+    const id1 = this.upQueue.events.subscribe(SyncQueueEventTypes.MutationEnqueued, (event) =>
+      cb(event.payload.queueSize)
     );
-    const id2 = this.upQueue.events.subscribe(
-      SyncQueueEventTypes.MutationDequeued,
-      (event) => cb(event.payload.queueSize)
+    const id2 = this.upQueue.events.subscribe(SyncQueueEventTypes.MutationDequeued, (event) =>
+      cb(event.payload.queueSize)
     );
     return () => {
       this.upQueue.events.unsubscribe(id1);
@@ -51,16 +219,232 @@ export class SpookySync<S extends SchemaStructure> {
     };
   }
 
+  // ---- Sync health -------------------------------------------------------
+  // `0` disables degraded reporting (config `syncHealth: false`). Resolved
+  // from config in Sp00kyClient and passed through the constructor options.
+  private readonly degradeAfterFailures: number;
+  /** Per-push RPC deadline; see {@link withPushTimeout}. */
+  private readonly pushTimeoutMs: number;
+  private consecutiveSyncFailures = 0;
+  private syncHealthStatus: SyncHealthStatus = 'healthy';
+  private lastSyncErrorKind: 'network' | 'application' | undefined;
+  private lastSyncErrorMessage: string | undefined;
+  // Latched `true` on the first successful sync round; never reset. Lets a UI
+  // tell a cold-start "connecting" phase (never reached the server) apart from
+  // a real lost connection after a working session.
+  private hasSyncedOnce = false;
+
+  // Self-heal: while degraded, re-drive sync on an exponential backoff so the
+  // app recovers on its own — even when the socket never actually dropped (in
+  // that case no `connected` event fires, so this re-registration is the ONLY
+  // thing that re-probes the server). Started on the degrade transition,
+  // cleared on recovery. Capped cadence so a long outage doesn't busy-loop.
+  private selfHealTimer: ReturnType<typeof setTimeout> | null = null;
+  private selfHealAttempts = 0;
+  private static readonly SELF_HEAL_BASE_MS = 2_000;
+  private static readonly SELF_HEAL_MAX_MS = 30_000;
+
+  /**
+   * Transport supervisor, when one was supplied. Sync only reads state from it;
+   * it never drives reconnects itself.
+   */
+  private readonly connectionSupervisor?: ConnectionSupervisor;
+  /**
+   * Mirror of the supervisor's state. Defaults to `connected` so a client
+   * constructed without a supervisor (tests, embedders) reports the same health
+   * shape it always has rather than a permanent false "disconnected".
+   */
+  private connectionState: ConnectionState = 'connected';
+
+  /** Current sync-health snapshot. */
+  get syncHealth(): SyncHealth {
+    return {
+      status: this.syncHealthStatus,
+      consecutiveFailures: this.consecutiveSyncFailures,
+      kind: this.syncHealthStatus === 'degraded' ? this.lastSyncErrorKind : undefined,
+      error: this.syncHealthStatus === 'degraded' ? this.lastSyncErrorMessage : undefined,
+      everConnected: this.hasSyncedOnce,
+      connection: this.connectionState,
+    };
+  }
+
+  /**
+   * Observe sync health. The callback fires immediately with the current
+   * status and again on every healthy↔degraded transition. Returns an
+   * unsubscribe. Mirrors {@link subscribeToPendingMutations}.
+   */
+  subscribeToSyncHealth(cb: (health: SyncHealth) => void): () => void {
+    cb(this.syncHealth);
+    const id = this.events.subscribe(SyncEventTypes.SyncHealthChanged, (event) =>
+      cb(event.payload)
+    );
+    return () => this.events.unsubscribe(id);
+  }
+
+  private emitSyncHealth(): void {
+    this.events.emit(SyncEventTypes.SyncHealthChanged, this.syncHealth);
+  }
+
+  /**
+   * Mirror the supervisor's transport state into {@link SyncHealth} and emit on
+   * every change, so a UI can react to a dropped socket immediately instead of
+   * waiting for `degradeAfterFailures` failed rounds. `status` is untouched:
+   * a brief reconnect is not a degradation.
+   *
+   * No explicit unsubscribe: the supervisor is owned by the same client and
+   * drops all subscribers in its own `dispose()`, which `Sp00kyClient.close()`
+   * calls first.
+   */
+  private subscribeToConnectionState(): void {
+    if (!this.connectionSupervisor) return;
+    this.connectionSupervisor.subscribe((state) => {
+      if (this.connectionState === state) return;
+      this.connectionState = state;
+      this.emitSyncHealth();
+    });
+  }
+
+  /**
+   * Fed by the scheduler once per drained sync round. Individual failures are
+   * absorbed by the queue's retry; only a run of `degradeAfterFailures`
+   * consecutive failures flips the status to `degraded`, and the next clean
+   * round flips it back. No-op when reporting is disabled (`degradeAfterFailures`
+   * is 0).
+   */
+  private recordSyncOutcome(ok: boolean, error?: unknown): void {
+    if (this.degradeAfterFailures <= 0) return;
+    if (ok) {
+      // Latch first-ever success so a UI can drop the connecting phase. Set
+      // before the early return so a clean cold start (0 prior failures) counts.
+      this.hasSyncedOnce = true;
+      if (this.consecutiveSyncFailures === 0) return;
+      this.consecutiveSyncFailures = 0;
+      if (this.syncHealthStatus !== 'healthy') {
+        this.syncHealthStatus = 'healthy';
+        this.lastSyncErrorKind = undefined;
+        this.lastSyncErrorMessage = undefined;
+        this.stopSelfHeal();
+        this.logger.info(
+          { Category: 'sp00ky-client::Sp00kySync::syncHealth' },
+          'Sync recovered; health back to healthy'
+        );
+        this.emitSyncHealth();
+      }
+      return;
+    }
+    this.consecutiveSyncFailures++;
+    this.lastSyncErrorKind = classifySyncError(error);
+    this.lastSyncErrorMessage = error instanceof Error ? error.message : String(error);
+    if (
+      this.syncHealthStatus !== 'degraded' &&
+      this.consecutiveSyncFailures >= this.degradeAfterFailures
+    ) {
+      this.syncHealthStatus = 'degraded';
+      this.logger.warn(
+        {
+          consecutiveFailures: this.consecutiveSyncFailures,
+          kind: this.lastSyncErrorKind,
+          error,
+          Category: 'sp00ky-client::Sp00kySync::syncHealth',
+        },
+        'Sync degraded after sustained failures'
+      );
+      this.emitSyncHealth();
+      this.startSelfHeal();
+    }
+  }
+
+  /**
+   * Begin self-heal retries (no-op if already running). Started on the
+   * healthy→degraded transition; {@link recordSyncOutcome} stops it on recovery.
+   */
+  private startSelfHeal(): void {
+    if (this.selfHealTimer !== null) return;
+    this.selfHealAttempts = 0;
+    this.scheduleSelfHeal();
+  }
+
+  private scheduleSelfHeal(): void {
+    const delay = Math.min(
+      Sp00kySync.SELF_HEAL_MAX_MS,
+      Sp00kySync.SELF_HEAL_BASE_MS * 2 ** this.selfHealAttempts
+    );
+    this.selfHealTimer = setTimeout(async () => {
+      this.selfHealTimer = null;
+      if (this.syncHealthStatus !== 'degraded') return;
+      this.selfHealAttempts++;
+      this.logger.debug(
+        {
+          attempt: this.selfHealAttempts,
+          delayMs: delay,
+          Category: 'sp00ky-client::Sp00kySync::selfHeal',
+        },
+        'Self-heal: re-driving sync while degraded'
+      );
+      try {
+        // Retry whatever is still queued first; the failing op (register or
+        // mutation) was re-queued by the queue, so this re-probes the server
+        // and reports the outcome through the scheduler → recordSyncOutcome.
+        if (this.upQueue.size > 0) {
+          await this.scheduler.syncUp();
+        } else if (this.downQueue.size > 0) {
+          await this.scheduler.syncDown();
+        } else {
+          // Nothing queued (e.g. the failing op was rolled back + dropped):
+          // re-register active queries — mirroring the reconnect handler — so
+          // there's a concrete op whose success flips health. If there are no
+          // active queries either, probe connectivity directly.
+          const hashes = this.dataModule.getActiveQueryHashes();
+          if (hashes.length > 0) {
+            for (const hash of hashes) {
+              this.scheduler.enqueueDownEvent({ type: 'register', payload: { hash } });
+            }
+            await this.scheduler.syncDown();
+          } else {
+            await this.remote.query('RETURN true');
+            this.recordSyncOutcome(true);
+          }
+        }
+      } catch (err) {
+        // Only the direct connectivity probe can throw here (syncUp/syncDown
+        // swallow + self-report); treat a probe failure as another failed round.
+        this.recordSyncOutcome(false, err);
+      }
+      // Keep retrying until recovery. recordSyncOutcome(true) calls stopSelfHeal
+      // (clearing any pending timer), so only continue while still degraded.
+      if (this.syncHealthStatus === 'degraded') this.scheduleSelfHeal();
+    }, delay);
+  }
+
+  private stopSelfHeal(): void {
+    if (this.selfHealTimer !== null) {
+      clearTimeout(this.selfHealTimer);
+      this.selfHealTimer = null;
+    }
+    this.selfHealAttempts = 0;
+  }
+
+  /**
+   * Release a deregistered query's remote view immediately instead of leaving
+   * it to the TTL sweep. Off by default; see the reasoning in
+   * {@link cleanupQuery}. Kept as a field rather than deleted so the eager path
+   * can be re-enabled in a test once the subquery-body repair path exists.
+   */
+  private readonly releaseQueriesEagerly = false;
+
   constructor(
-    private local: LocalDatabaseService,
+    private local: LocalStore,
     private remote: RemoteDatabaseService,
     private cache: CacheModule,
     private dataModule: DataModule<S>,
     private schema: S,
-    logger: Logger
+    logger: Logger,
+    options?: Sp00kySyncOptions
   ) {
-    this.logger = logger.child({ service: 'SpookySync' });
-    this.upQueue = new UpQueue(this.local, this.logger);
+    this.logger = logger.child({ service: 'Sp00kySync' });
+    this.upQueue = new UpQueue(this.local, this.logger, (dropped) =>
+      this.onMutationDropped(dropped)
+    );
     this.downQueue = new DownQueue(this.local, this.logger);
     this.syncEngine = new SyncEngine(this.remote, this.cache, this.schema, this.logger);
     this.scheduler = new SyncScheduler(
@@ -69,43 +453,685 @@ export class SpookySync<S extends SchemaStructure> {
       this.processUpEvent.bind(this),
       this.processDownEvent.bind(this),
       this.logger,
-      this.handleRollback.bind(this)
+      this.handleRollback.bind(this),
+      this.recordSyncOutcome.bind(this),
+      this.handleMutationSettled.bind(this)
     );
+    this.refSyncIntervalMs = resolveListRefPollInterval(options?.refSyncIntervalMs);
+    this.anonLiveEnabled = options?.anonymousLiveQueries ?? false;
+    this.degradeAfterFailures = Math.max(0, options?.degradeAfterConsecutiveFailures ?? 3);
+    this.pushTimeoutMs = Math.max(0, options?.pushTimeoutMs ?? 30_000);
+    this.connectionSupervisor = options?.connectionSupervisor;
   }
 
   /**
    * Initializes the synchronization system.
    * Starts the scheduler and initiates the initial sync cycles.
-   * @param clientId The unique identifier for this client instance.
    * @throws Error if already initialized.
    */
-  public async init(clientId: string) {
-    if (this.isInit) throw new Error('SpookySync is already initialized');
-    this.clientId = clientId;
+  public async init() {
+    if (this.isInit) throw new Error('Sp00kySync is already initialized');
     this.isInit = true;
-    await this.scheduler.init();
-    void this.scheduler.syncUp();
+    await this.scheduler.init({ loadOutbox: this.tabRole !== 'follower' });
+    this.subscribeToReconnect();
+    this.subscribeToConnectionState();
     void this.scheduler.syncUp();
     void this.scheduler.syncDown();
-    void this.startRefLiveQueries();
+    // No initial LIVE subscription — wait for `setCurrentUserId` to fire
+    // from the auth subscription. In dedicated mode the table name
+    // depends on the authenticated user, and an unauthenticated
+    // subscription wouldn't match any of the per-user tables anyway.
+    //
+    // Exception: when anonymous live queries are enabled, start realtime now
+    // against the shared `_00_list_ref_anon` table so a logged-out client
+    // syncs immediately. `setCurrentUserId` re-points LIVE to the per-user
+    // table on sign-in. Guard on `currentUserId` because the auth callback can
+    // fire (and authenticate) before `init()` runs — don't clobber that back
+    // to the anon table. `setCurrentUserId(null)` is a no-op on first load
+    // (it's already null), so this is the only place anon realtime starts.
+    if (this.anonLiveEnabled && !this.currentUserId) {
+      this.startListRefPoll();
+      this.restartRefLiveQuery().catch((err) => {
+        this.logger.debug(
+          { err, Category: 'sp00ky-client::Sp00kySync::init' },
+          'Anonymous ref LIVE start failed; relying on periodic poll fallback'
+        );
+      });
+    }
+  }
+
+  // ---- shared-tabs roles ------------------------------------------------------
+
+  /** Set BEFORE init(): shapes what init boots (a follower loads no outbox and
+   *  never starts LIVE; its own registration/poll paths stay untouched). */
+  public setTabContext(role: 'solo' | 'leader' | 'follower', tabId: string | null): void {
+    this.tabRole = role;
+    this.tabId = tabId;
+  }
+
+  /** Leader duties: drain the shared outbox, own the single list_ref LIVE,
+   *  relay LIVE events and rollbacks to followers via `hub`. Idempotent for a
+   *  boot-time leader; a runtime promotion (failover) reloads the outbox,
+   *  which now holds EVERY tab's rows, and restarts LIVE under this session. */
+  public async promoteToLeader(hub: LeaderSyncHub): Promise<void> {
+    this.tabRole = 'leader';
+    this.hub = hub;
+    this.forwarder = null;
+    hub.onFollowerMessage = (tabId, msg) => {
+      void tabId;
+      switch (msg.type) {
+        case 'sync-hello':
+          break;
+        case 'mutation-enqueued':
+          void this.enqueueForwardedMutation(msg.mutationId);
+          break;
+        case 'request-poll':
+          this.listRefIdleStreak = 0;
+          break;
+      }
+    };
+    if (this.isInit) {
+      await this.upQueue.loadFromDatabase();
+      void this.scheduler.syncUp();
+      if (this.currentUserId || this.anonLiveEnabled) {
+        this.startListRefPoll();
+        await this.restartRefLiveQuery().catch((err) => {
+          this.logger.warn(
+            { err, Category: 'sp00ky-client::Sp00kySync::promoteToLeader' },
+            'LIVE restart failed on promotion; poll fallback covers it'
+          );
+        });
+      }
+    }
+  }
+
+  /** Follower duties: no outbox drain, no LIVE. Mutations forward to the
+   *  leader; everything else (registration, per-query sync, poll) runs
+   *  against this tab's own remote session as usual. */
+  public demoteToFollower(forwarder: SyncForwarder): void {
+    this.tabRole = 'follower';
+    this.hub = null;
+    this.forwarder = forwarder;
+    void this.killRefLiveQuery();
+    forwarder.onLeaderMessage = (msg) => {
+      switch (msg.type) {
+        case 'list-ref-change':
+          void this.applyRelayedListRefChange(msg).catch((err) => {
+            this.logger.error(
+              { err, Category: 'sp00ky-client::Sp00kySync::relay' },
+              'Relayed list_ref change failed'
+            );
+          });
+          break;
+        case 'mutation-rolled-back':
+          this.events.emit(SyncEventTypes.MutationRolledBack, {
+            eventType: msg.eventType,
+            recordId: msg.recordId,
+            error: msg.error,
+          });
+          break;
+        default:
+          // db-ready and ingest-relay are handled by the engine/cache wiring
+          // in Sp00kyClient before messages reach this handler.
+          break;
+      }
+    };
+  }
+
+  /**
+   * A pending mutation was discarded because it can never be sent.
+   *
+   * This is a lost write, so it must not stay invisible. Every failure in this
+   * chain used to be a `logger.error` an app running `logLevel: 'fatal'` never
+   * shows, which is how an outbox could sit undrained for hours with the UI
+   * reporting nothing. Surfaces as a rollback event (the mutation will never
+   * apply, which is what a subscriber needs to know) and degrades sync health.
+   */
+  private onMutationDropped(dropped: {
+    mutationId: string;
+    recordId?: string;
+    mutationType?: string;
+    reason: string;
+  }): void {
+    this.logger.error(
+      { ...dropped, Category: 'sp00ky-client::Sp00kySync::onMutationDropped' },
+      'Dropped a pending mutation that can never be sent'
+    );
+    this.recordSyncOutcome(false, new Error(`dropped mutation: ${dropped.reason}`));
+    this.events.emit(SyncEventTypes.MutationRolledBack, {
+      eventType: (dropped.mutationType as 'create' | 'update' | 'delete') ?? 'update',
+      recordId: dropped.recordId ?? dropped.mutationId,
+      error: `dropped: ${dropped.reason}`,
+    });
+  }
+
+  /** A forwarded outbox row from a follower: load + drain it. Idempotent. */
+  public async enqueueForwardedMutation(mutationId: string): Promise<void> {
+    if (this.tabRole !== 'leader') return;
+    await this.upQueue.enqueueFromDatabase(mutationId);
+  }
+
+  /** A relayed `_00_list_ref` LIVE event: resolve against THIS tab's queries
+   *  and run the exact same handling the LIVE subscription would have. */
+  private async applyRelayedListRefChange(msg: {
+    action: 'CREATE' | 'UPDATE' | 'DELETE';
+    queryId: string;
+    recordId: string;
+    version: number;
+    parent: boolean;
+  }): Promise<void> {
+    const queryId = parseRecordIdString(msg.queryId);
+    // Foreign query (another tab's session-salted hash): not ours, ignore.
+    if (!this.dataModule.getQueryById(queryId)) return;
+    const recordId = parseRecordIdString(msg.recordId);
+    if (msg.parent) {
+      await this.handleRemoteSubqueryChange(msg.action, queryId, recordId, msg.version);
+    } else {
+      await this.handleRemoteListRefChange(msg.action, queryId, recordId, msg.version);
+    }
+  }
+
+  /** One immediate poll cycle (failover convergence). */
+  public async forcePollRound(): Promise<void> {
+    this.listRefIdleStreak = 0;
+    await this.pollListRefForActiveQueries().catch(() => false);
+  }
+
+  /**
+   * Quiesce all sync activity ahead of a local-bucket switch. After this
+   * resolves, nothing in the sync module writes to the local store: the poll
+   * loop is stopped AND its in-flight tick awaited, LIVE is killed, debounce
+   * timers are cancelled (their outbox rows are already persisted), and the
+   * scheduler has drained its in-flight queue item — including that item's
+   * outbox-row delete, which must land in the OLD bucket. Queued down-events
+   * are dropped (they reference old-bucket query rows; the post-switch rebind
+   * re-enqueues registrations). The old user's un-pushed outbox is deliberately
+   * NOT drained: the remote session already belongs to the next user.
+   */
+  public async prepareBucketSwitch(): Promise<void> {
+    this.stopSelfHeal();
+    this.stopListRefPoll();
+    if (this.listRefPollInFlight) await this.listRefPollInFlight;
+    await this.killRefLiveQuery();
+    this.upQueue.clearDebounceTimers();
+    await this.scheduler.pause();
+    this.downQueue.clear();
+    this.stillRemoteStreaks.clear();
+    this.logger.info(
+      { Category: 'sp00ky-client::Sp00kySync::prepareBucketSwitch' },
+      'Sync quiesced for bucket switch'
+    );
+  }
+
+  /**
+   * Resume syncing against the freshly-opened bucket: reload the mutation
+   * outbox from ITS `_00_pending_mutations` (the new user's own un-pushed
+   * offline work) and restart the scheduler. LIVE + the list_ref poll restart
+   * via the `setCurrentUserId` call that follows in the auth listener.
+   */
+  public async completeBucketSwitch(): Promise<void> {
+    await this.upQueue.loadFromDatabase();
+    this.scheduler.resume();
+    this.logger.info(
+      { Category: 'sp00ky-client::Sp00kySync::completeBucketSwitch' },
+      'Sync resumed after bucket switch'
+    );
+  }
+
+  /**
+   * Push the authenticated user's record id from the parent client's
+   * auth subscription. Tears down the existing `_00_list_ref` LIVE (if
+   * any) and re-registers it under the new user's dedicated table so
+   * SurrealDB binds the permission rule under the post-flip auth
+   * context. Pass `null` on sign-out.
+   *
+   * The dedicated `_00_list_ref_user_<id>` table is created lazily by
+   * the SSP when the first query registration arrives, which may be
+   * concurrent with this call. We retry the LIVE registration with a
+   * short backoff so a "table not found" race resolves without
+   * surfacing as a permanent auth-loading hang.
+   */
+  public async setCurrentUserId(userId: string | null): Promise<void> {
+    if (this.currentUserId === userId) return;
+    this.currentUserId = userId;
+    if (!userId) {
+      if (this.anonLiveEnabled) {
+        // Signed out but anonymous realtime is on: keep the poll running and
+        // re-point LIVE from the (now stale) per-user table to the shared
+        // `_00_list_ref_anon`. `startListRefPoll` is idempotent; the poll
+        // re-resolves `listRefTable()` each tick so it follows automatically.
+        this.startListRefPoll();
+        await this.restartRefLiveQuery().catch((err) => {
+          this.logger.debug(
+            { err, Category: 'sp00ky-client::Sp00kySync::setCurrentUserId' },
+            'Anonymous ref LIVE restart failed; relying on periodic poll fallback'
+          );
+        });
+        return;
+      }
+      await this.killRefLiveQuery();
+      this.stopListRefPoll();
+      return;
+    }
+    // Start periodic polling FIRST so we have a deterministic fallback
+    // even when LIVE registration fails or SurrealDB drops a delivery.
+    this.startListRefPoll();
+    // Try to start LIVE with backoff for low-latency delivery on the
+    // happy path; the poll handles the rest.
+    const attemptDelays = [0, 250, 500, 1000, 2000];
+    for (let i = 0; i < attemptDelays.length; i++) {
+      if (attemptDelays[i] > 0) {
+        this._liveRetryCount++;
+        await new Promise((r) => setTimeout(r, attemptDelays[i]));
+      }
+      try {
+        await this.restartRefLiveQuery();
+        return;
+      } catch (err) {
+        this.logger.debug(
+          { err, attempt: i + 1, Category: 'sp00ky-client::Sp00kySync::setCurrentUserId' },
+          'Ref LIVE start failed; relying on periodic poll fallback'
+        );
+      }
+    }
+  }
+
+  private startListRefPoll(): void {
+    if (this.listRefPollRunning) return;
+    this.listRefPollRunning = true;
+    this.logger.debug(
+      {
+        intervalMs: this.refSyncIntervalMs,
+        Category: 'sp00ky-client::Sp00kySync::startListRefPoll',
+      },
+      'list_ref poll loop started'
+    );
+    const schedule = (delayMs: number) => {
+      this.listRefPollTimer = setTimeout(async () => {
+        if (!this.listRefPollRunning) return;
+        let changed = false;
+        const tick = (async () => {
+          changed = await this.pollListRefForActiveQueries();
+        })();
+        this.listRefPollInFlight = tick.catch(() => {});
+        try {
+          await tick;
+        } finally {
+          this.listRefPollInFlight = null;
+          if (!this.listRefPollRunning) return;
+          // Reset the idle streak on any observed change so the poll snaps
+          // back to the fast base cadence; otherwise grow it so a quiet page
+          // backs off toward the cap. (`handleRemoteListRefChange` also resets
+          // it when a LIVE event lands.)
+          this.listRefIdleStreak = changed ? 0 : this.listRefIdleStreak + 1;
+          const next = listRefPollDelayMs({
+            idleStreak: this.listRefIdleStreak,
+            baseIntervalMs: this.refSyncIntervalMs,
+          });
+          schedule(next);
+        }
+      }, delayMs);
+    };
+    schedule(this.refSyncIntervalMs);
+  }
+
+  private stopListRefPoll(): void {
+    this.listRefPollRunning = false;
+    if (this.listRefPollTimer !== null) {
+      clearTimeout(this.listRefPollTimer);
+      this.listRefPollTimer = null;
+    }
+  }
+
+  /**
+   * One poll cycle: refetch `_00_list_ref` for every active query. Returns
+   * whether ANY query's remoteArray actually changed — the scheduler uses this
+   * to drive the adaptive idle backoff.
+   *
+   * Also the ONLY health signal that runs while the page is idle. Sync health is
+   * otherwise activity-driven (mutations/registrations via the scheduler,
+   * reconnect re-registration, self-heal), so on a quiet page a stale `degraded`
+   * would linger until the next mutation and a genuine idle drop would be
+   * invisible. We fold the cycle's aggregate reachability into `recordSyncOutcome`
+   * so idle health self-recovers (and self-degrades) with no user action. A clean
+   * cycle is idempotent when already healthy (`recordSyncOutcome` early-returns at
+   * `consecutiveSyncFailures === 0`), so a healthy idle page pays nothing.
+   */
+  private async pollListRefForActiveQueries(): Promise<boolean> {
+    const hashes = this.dataModule.getActiveQueryHashes();
+    if (hashes.length === 0) {
+      // No active queries to piggyback on, but health still needs a heartbeat —
+      // probe connectivity directly so an idle page with no live queries doesn't
+      // go blind. Cheap, and gated by the same adaptive backoff (≤5s idle cap).
+      try {
+        await this.remote.query('RETURN true');
+        this.recordSyncOutcome(true);
+      } catch (err) {
+        this.recordSyncOutcome(false, err);
+      }
+      return false;
+    }
+    let anyChanged = false;
+    // `reached` = the server answered at least once this cycle (a success, or an
+    // *application* error, which still proves reachability). `firstNetworkErr`
+    // holds the first network-classified failure. A cycle that only produced
+    // network errors reports the outcome as a down round; a mixed/app cycle counts
+    // as reached; an all-application cycle reports nothing (that's a query-shape
+    // fault owned by the registration path, not a reachability signal).
+    let reached = false;
+    let firstNetworkErr: unknown;
+    for (const hash of hashes) {
+      try {
+        if (await this.refetchListRefForQuery(hash)) anyChanged = true;
+        reached = true;
+      } catch (err) {
+        if (classifySyncError(err) === 'network') {
+          if (firstNetworkErr === undefined) firstNetworkErr = err;
+        } else {
+          reached = true;
+        }
+        this.logger.debug(
+          {
+            err: (err as Error)?.message ?? err,
+            hash,
+            Category: 'sp00ky-client::Sp00kySync::pollListRefForActiveQueries',
+          },
+          'Per-query list_ref poll failed'
+        );
+      }
+    }
+    // Call the private outcome recorder directly rather than routing through the
+    // scheduler — the scheduler only reports on rounds that drained ≥1 queue item
+    // (`processedAny`), and this isn't a queue round.
+    if (reached) {
+      this.recordSyncOutcome(true);
+    } else if (firstNetworkErr !== undefined) {
+      this.recordSyncOutcome(false, firstNetworkErr);
+    }
+    return anyChanged;
+  }
+
+  /**
+   * Pull the upstream list_ref entries for `queryHash`, diff them
+   * against the local `remoteArray` cache, sync any added/updated rows
+   * through the SyncEngine, then persist the new remoteArray. This is
+   * the same shape `createRemoteQuery` does for its initial fetch and
+   * what `handleRemoteListRefChange` does per-LIVE-event — we reuse
+   * it on a timer as a fallback for missed LIVE notifications.
+   */
+  private async refetchListRefForQuery(queryHash: string): Promise<boolean> {
+    const queryState = this.dataModule.getQueryByHash(queryHash);
+    if (!queryState) return false;
+    const listRefTbl = this.listRefTable();
+    const [items, serverRowCount] = await this.remote.query<
+      [{ out: RecordId<string>; version: number }[], number | null]
+    >(`${buildListRefSelect(listRefTbl)};\n${buildQueryRowCountSelect()}`, {
+      in: queryState.config.id,
+    });
+    if (!Array.isArray(items)) return false;
+    const fresh: RecordVersionArray = items.map((item) => [encodeRecordId(item.out), item.version]);
+    // Capture which ids LEFT the query's window (present in the cached
+    // remoteArray, absent from `fresh`) BEFORE we overwrite remoteArray — these
+    // are cross-window deletes (or rows that scrolled out). They drive the
+    // forced re-render below.
+    const prevRemote = queryState.config.remoteArray ?? [];
+    const freshIds = new Set(fresh.map(([id]) => id));
+    const removedIds = prevRemote.filter(([id]) => !freshIds.has(id)).map(([id]) => id);
+    // Idempotent poll: only persist the remoteArray when it actually changed.
+    // The poll runs continuously as a LIVE fallback, so on a quiet page `fresh`
+    // equals the cached array every tick — re-writing it (an `UPDATE _00_query`
+    // each cycle, per active query) was pure churn and the bulk of the idle
+    // traffic. `recordVersionArraysEqual` is order-insensitive because the
+    // list_ref SELECT has no `ORDER BY`.
+    const changed = !recordVersionArraysEqual(fresh, queryState.config.remoteArray);
+    if (changed) {
+      // Update the cached remoteArray so the next diff/sync sees the new state.
+      // `syncQuery` (below) then writes through `cache.saveBatch`, which UPSERTs
+      // the local DB row and ingests it into the in-browser SSP — the SSP's
+      // stream updates run `processStreamUpdate`, which re-queries the local DB
+      // and notifies subscribers. We skip an explicit `notifyQuerySynced`
+      // because that path races the stream-update path (can notify with stale
+      // records).
+      await this.dataModule.updateQueryRemoteArray(queryHash, fresh, { serverRowCount });
+    }
+    // Run `syncQuery` every tick regardless: it's a no-op when localArray has
+    // caught up to remoteArray (`if (!diff) return`, issues no query), but it
+    // covers the rare case where remoteArray is stable yet localArray is behind
+    // (a prior record fetch failed) — so a missed row still gets retried.
+    // For REMOVALS it runs the ids through `handleRemovedRecords`, which deletes
+    // confirmed-gone records from the local DB.
+    try {
+      await this.syncQuery(queryHash);
+    } catch (err) {
+      this.logger.info(
+        {
+          err: (err as Error)?.message ?? err,
+          queryHash,
+          Category: 'sp00ky-client::Sp00kySync::refetchListRefForQuery',
+        },
+        'syncQuery failed during poll'
+      );
+    }
+    // Cross-session fallback for `.related()` child rows: the LIVE-permission
+    // gap can drop child-edge notifications, so converge their bodies on the
+    // poll too (idempotent — no-op when nothing changed).
+    await this.syncSubqueryChildren(queryHash).catch((err) => {
+      this.logger.info(
+        {
+          err: (err as Error)?.message ?? err,
+          queryHash,
+          Category: 'sp00ky-client::Sp00kySync::refetchListRefForQuery',
+        },
+        'Subquery child sync failed during poll'
+      );
+    });
+    // A REMOVAL needs no record fetch, so unlike the added-row path it doesn't
+    // get a re-render from the SSP stream on this code path reliably (and the
+    // non-windowed window-0 query re-queries the local DB rather than the id-set).
+    // Force a re-materialize + notify so the deleted row drops from the list in
+    // this (second) window — the reliable, LIVE-independent cross-window path.
+    if (removedIds.length > 0) {
+      try {
+        await this.dataModule.notifyQuerySynced(queryHash);
+      } catch (err) {
+        this.logger.info(
+          {
+            err: (err as Error)?.message ?? err,
+            queryHash,
+            Category: 'sp00ky-client::Sp00kySync::refetchListRefForQuery',
+          },
+          'notifyQuerySynced failed during poll-removal re-render'
+        );
+      }
+    }
+    return changed;
+  }
+
+  /**
+   * Resolve the current `_00_list_ref` table name for the active auth
+   * context. Public so the `createRemoteQuery` initial-fetch path can
+   * read from the right per-user table.
+   *
+   * Reads the user id from `DataModule` rather than the local mirror,
+   * because `DataModule.setCurrentUserId` runs synchronously from the
+   * auth callback (before any `await`), whereas `sync.setCurrentUserId`
+   * is async — the userQuery's initial fetch can fire between those
+   * two points and we need the correct table name immediately.
+   */
+  public listRefTable(): string {
+    const userId = this.dataModule.getCurrentUserId();
+    // Unauthenticated with the flag on → the shared `_00_list_ref_anon` table.
+    if (userId == null && this.anonLiveEnabled) {
+      return listRefTableFor(this.refMode, ANON_USER_ID);
+    }
+    return listRefTableFor(this.refMode, userId);
+  }
+
+  private async killRefLiveQuery(): Promise<void> {
+    if (this.liveQueryUnsubscribe) {
+      try {
+        this.liveQueryUnsubscribe();
+      } catch {
+        /* ignore */
+      }
+      this.liveQueryUnsubscribe = null;
+    }
+    if (this.currentLiveQueryUuid !== null) {
+      // A LIVE subscription is scoped to its WebSocket session, so after a
+      // reconnect the server-side one is already gone and there is nothing to
+      // KILL. Sending it anyway either fails (no connection) or races the fresh
+      // socket's readiness while holding up the restart behind it. Local
+      // bookkeeping is cleared either way.
+      if (this.remote.getStatus() === 'connected') {
+        try {
+          await this.remote.query('KILL $u', { u: this.currentLiveQueryUuid });
+        } catch (err) {
+          this.logger.debug(
+            { err, Category: 'sp00ky-client::Sp00kySync::killRefLiveQuery' },
+            'Prior LIVE KILL failed; continuing'
+          );
+        }
+      }
+      this.currentLiveQueryUuid = null;
+    }
+  }
+
+  private async restartRefLiveQuery(): Promise<void> {
+    await this.killRefLiveQuery();
+    await this.startRefLiveQueries();
+  }
+
+  /**
+   * Drop local LIVE bookkeeping without issuing a `KILL`.
+   *
+   * Called when the socket dies. The server-side subscription is scoped to that
+   * WebSocket session and died with it, so there is nothing left to kill — and
+   * by the time the reconnect handler runs, the client reports `connected`
+   * again, which would otherwise send a `KILL` for a stale uuid on the *new*
+   * session and hold up the restart queued behind it.
+   */
+  private invalidateRefLiveQuery(): void {
+    if (this.liveQueryUnsubscribe) {
+      try {
+        this.liveQueryUnsubscribe();
+      } catch {
+        /* ignore */
+      }
+      this.liveQueryUnsubscribe = null;
+    }
+    this.currentLiveQueryUuid = null;
+  }
+
+  // Only the connect that follows a prior drop counts as a reconnect; the
+  // initial connect after init() must not trigger a refetch storm.
+  //
+  // Both drop events have to be watched. The SDK publishes `disconnected` ONLY
+  // when it has given up entirely (attempts exhausted, or the engine
+  // terminated); an ordinary recovered drop goes `error` -> `reconnecting` ->
+  // `connected` and never touches `disconnected`. Listening for `disconnected`
+  // alone therefore misses every successful reconnect — the exact case this
+  // handler exists for — and leaves the dead server-side LIVE in place with the
+  // poll as the only sync path.
+  private subscribeToReconnect() {
+    const client = this.remote.getClient();
+    client.subscribe('disconnected', () => {
+      this.needsResubscribe = true;
+      this.invalidateRefLiveQuery();
+      this.logger.info(
+        { Category: 'sp00ky-client::Sp00kySync::onDisconnect' },
+        'Remote disconnected'
+      );
+    });
+    client.subscribe('reconnecting', () => {
+      this.needsResubscribe = true;
+      this.invalidateRefLiveQuery();
+      this.logger.info(
+        { Category: 'sp00ky-client::Sp00kySync::onReconnecting' },
+        'Remote socket dropped; awaiting reconnect'
+      );
+    });
+    client.subscribe('connected', () => {
+      if (!this.needsResubscribe) return;
+      this.needsResubscribe = false;
+      // A flapping socket produces reconnecting -> connected repeatedly, and
+      // each cycle used to re-register EVERY active query (a busy app has
+      // dozens). That is the "everything reloads about a second after a blip"
+      // symptom: the SDK's retryDelay is 1s, so the refetch lands right after
+      // the drop the user never saw. Collapse bursts into one refetch.
+      const sinceLast = Date.now() - this.lastReconnectRefetchAt;
+      if (sinceLast < Sp00kySync.RECONNECT_REFETCH_COOLDOWN_MS) {
+        this.logger.debug(
+          { sinceLast, Category: 'sp00ky-client::Sp00kySync::onReconnect' },
+          'Reconnected again within the cooldown; skipping duplicate refetch'
+        );
+        return;
+      }
+      this.lastReconnectRefetchAt = Date.now();
+      const hashes = this.dataModule.getActiveQueryHashes();
+      this.logger.info(
+        { queries: hashes.length, Category: 'sp00ky-client::Sp00kySync::onReconnect' },
+        'Remote reconnected, refetching active queries'
+      );
+      for (const hash of hashes) {
+        this.scheduler.enqueueDownEvent({ type: 'register', payload: { hash } });
+      }
+      // The WS reconnect leaves the server-side LIVE subscription dead — the
+      // re-enqueued `register` events only re-fetch initial state, they don't
+      // re-subscribe. Without this, LIVE never recovers after a reconnect and
+      // the poll silently becomes the sole sync path (and never backs off).
+      // Authenticated → per-user table; signed-out with anon live enabled →
+      // the shared `_00_list_ref_anon`. Otherwise there's no table to re-bind.
+      if (this.currentUserId || this.anonLiveEnabled) {
+        this.restartRefLiveQuery().catch((err) => {
+          this.logger.debug(
+            { err, Category: 'sp00ky-client::Sp00kySync::onReconnect' },
+            'LIVE restart after reconnect failed; relying on poll fallback'
+          );
+        });
+      }
+    });
   }
 
   private async startRefLiveQueries() {
+    // Shared-tabs follower: exactly one LIVE per user exists, on the leader;
+    // its events reach this tab through the relay.
+    if (this.tabRole === 'follower') return;
+    const tableName = this.listRefTable();
     this.logger.debug(
-      { clientId: this.clientId, Category: 'spooky-client::SpookySync::startRefLiveQueries' },
+      { tableName, Category: 'sp00ky-client::Sp00kySync::startRefLiveQueries' },
       'Starting ref live queries'
     );
 
-    const [queryUuid] = await this.remote.query<[Uuid]>(
-      'LIVE SELECT * FROM _spooky_list_ref'
-    );
+    const [queryUuid] = await this.remote.query<[Uuid]>(`LIVE SELECT * FROM ${tableName}`);
+    this.currentLiveQueryUuid = queryUuid;
 
-    (await this.remote.getClient().liveOf(queryUuid)).subscribe((message) => {
+    const live = await this.remote.getClient().liveOf(queryUuid);
+    this.liveQueryUnsubscribe = live.subscribe((message) => {
       this.logger.debug(
-        { message, Category: 'spooky-client::SpookySync::startRefLiveQueries' },
+        { message, Category: 'sp00ky-client::Sp00kySync::startRefLiveQueries' },
         'Live update received'
       );
       if (message.action === 'KILLED') return;
+      // Subquery child edges (rows with `parent` set) are NOT primary window
+      // rows — the client's `RecordVersionArray` only tracks primary rows, so
+      // routing them through `handleRemoteListRefChange` would surface them as
+      // spurious "added" diffs and pollute the window. Instead route them to
+      // the dedicated child-body sync so `.related()` data stays realtime
+      // cross-session (the LIVE-permission gap otherwise leaves it to the poll).
+      if ((message.value as { parent?: unknown }).parent != null) {
+        this.handleRemoteSubqueryChange(
+          message.action,
+          message.value.in as RecordId<string>,
+          message.value.out as RecordId<string>,
+          message.value.version as number
+        ).catch((err) => {
+          this.logger.error(
+            { err, Category: 'sp00ky-client::Sp00kySync::startRefLiveQueries' },
+            'Error handling remote subquery change'
+          );
+        });
+        return;
+      }
       this.handleRemoteListRefChange(
         message.action,
         message.value.in as RecordId<string>,
@@ -113,7 +1139,7 @@ export class SpookySync<S extends SchemaStructure> {
         message.value.version as number
       ).catch((err) => {
         this.logger.error(
-          { err, Category: 'spooky-client::SpookySync::startRefLiveQueries' },
+          { err, Category: 'sp00ky-client::Sp00kySync::startRefLiveQueries' },
           'Error handling remote list ref change'
         );
       });
@@ -126,16 +1152,47 @@ export class SpookySync<S extends SchemaStructure> {
     recordId: RecordId,
     version: number
   ) {
+    // Any LIVE delivery is evidence of activity — a CREATE/UPDATE/DELETE on a
+    // query's window, or a notification for an unknown local query. Reset the
+    // poll's idle streak so it snaps back to the fast base cadence (the page
+    // is clearly not idle), and record the timestamp as a liveness diagnostic.
+    this.lastLiveEventAt = Date.now();
+    this.listRefIdleStreak = 0;
+
+    // Shared-tabs leader: the list_ref table is USER-scoped, so this LIVE also
+    // carries events for FOLLOWER tabs' queries (their own session-salted
+    // hashes). Relay every primary event; each follower resolves the queryId
+    // against its own DataModule and ignores foreign ones. Then continue with
+    // this tab's own handling below.
+    this.hub?.broadcast({
+      type: 'list-ref-change',
+      action,
+      queryId: encodeRecordId(queryId),
+      recordId: encodeRecordId(recordId),
+      version,
+      parent: false,
+    });
+
+    // NOTE: DELETE is handled like CREATE/UPDATE below. When another window (or
+    // this one) deletes a record, the server's SSP removes it from `_00_list_ref`
+    // and the LIVE subscription delivers a DELETE here — `createDiffFromDbOp`
+    // turns it into a `removed: [recordId]` diff so the row drops from the window
+    // in realtime. (It was previously ignored, so other windows only caught up on
+    // reload / the slow poll.)
     const existing = this.dataModule.getQueryById(queryId);
 
     if (!existing) {
-      this.logger.warn(
-        {
-          queryId: queryId.toString(),
-          Category: 'spooky-client::SpookySync::handleRemoteListRefChange',
-        },
-        'Received remote update for unknown local query'
-      );
+      // With a hub attached, an unknown query is the NORMAL case (it belongs
+      // to a follower tab); without one it still warrants the warning.
+      if (!this.hub) {
+        this.logger.warn(
+          {
+            queryId: queryId.toString(),
+            Category: 'sp00ky-client::Sp00kySync::handleRemoteListRefChange',
+          },
+          'Received remote update for unknown local query'
+        );
+      }
       return;
     }
 
@@ -148,12 +1205,87 @@ export class SpookySync<S extends SchemaStructure> {
         recordId,
         version,
         localArray,
-        Category: 'spooky-client::SpookySync::handleRemoteListRefChange',
+        Category: 'sp00ky-client::Sp00kySync::handleRemoteListRefChange',
       },
       'Live update is being processed'
     );
     const diff = createDiffFromDbOp(action, recordId, version, localArray);
-    await this.syncEngine.syncRecords(diff);
+    // `config.id` is `_00_query:<hash>`, so its id-part IS the query hash
+    // (a SHA-256 over query content + sessionId) — the key DataModule uses.
+    const hash = extractIdPart(existing.config.id);
+
+    // Apply the event to `remoteArray` — the authoritative membership rows are
+    // now rendered FROM. Only registration and the poll used to write it, so a
+    // LIVE removal left the departed id in the list (in memory and persisted)
+    // until the next poll tick, which is up to 5s of showing a deleted row. This
+    // also persists the durable `_00_window` mirror, so the removal survives a
+    // reload with no network.
+    if (existing.config.membershipKnown && (diff.removed.length || diff.added.length)) {
+      const next = applyRecordVersionDiff(existing.config.remoteArray ?? [], diff);
+      if (!recordVersionArraysEqual(next, existing.config.remoteArray ?? [])) {
+        await this.dataModule.updateQueryRemoteArray(hash, next);
+      }
+    }
+
+    await this.runSyncForQuery(hash, diff);
+
+    // A removal-only diff sets `fetching` false in `runSyncForQuery`, so it gets
+    // no `flushPendingStreamUpdate`/`endFetching` re-render — and a removal needs
+    // no record fetch to trigger one either. Force it, mirroring what the poll
+    // path already does for its own removals (`refetchListRefForQuery`).
+    if (diff.removed.length > 0 && diff.added.length === 0 && diff.updated.length === 0) {
+      await this.dataModule.notifyQuerySynced(hash);
+    }
+  }
+
+  /**
+   * Handle a LIVE change to a SUBQUERY child edge (a `_00_list_ref` row with
+   * `parent` set) for a `.related()` query. Unlike primary rows, child rows
+   * must NOT touch the query's `localArray`/`remoteArray`/`rowCount`; we only
+   * keep the child BODY fresh in the local cache so the in-browser SSP's
+   * subquery-table dependency re-materializes the parent view.
+   *
+   * CREATE/UPDATE fetch+upsert the child body. DELETE is intentionally a
+   * no-op: a child leaving this query's set must not delete a body another
+   * query may still show (see `syncSubqueryChildren` deletion-safety note);
+   * a genuine record delete propagates via the normal delete path.
+   */
+  private async handleRemoteSubqueryChange(
+    action: 'CREATE' | 'UPDATE' | 'DELETE',
+    queryId: RecordId,
+    childId: RecordId,
+    version: number
+  ) {
+    this.lastLiveEventAt = Date.now();
+    this.listRefIdleStreak = 0;
+
+    if (action === 'DELETE') return;
+
+    // Relay child-edge events too (see handleRemoteListRefChange).
+    this.hub?.broadcast({
+      type: 'list-ref-change',
+      action,
+      queryId: encodeRecordId(queryId),
+      recordId: encodeRecordId(childId),
+      version,
+      parent: true,
+    });
+
+    const existing = this.dataModule.getQueryById(queryId);
+    if (!existing) return;
+
+    const item = { id: childId, version };
+    await this.syncEngine.syncRecords(
+      action === 'CREATE'
+        ? { added: [item], updated: [], removed: [] }
+        : { added: [], updated: [item], removed: [] }
+    );
+
+    // Keep the in-memory child array in step so the poll's idempotent diff
+    // doesn't re-fetch this body on the next tick.
+    const key = encodeRecordId(childId);
+    const prev = existing.config.subqueryRemoteArray ?? [];
+    existing.config.subqueryRemoteArray = [...prev.filter(([id]) => id !== key), [key, version]];
   }
 
   /**
@@ -164,50 +1296,92 @@ export class SpookySync<S extends SchemaStructure> {
     this.scheduler.enqueueDownEvent(event);
   }
 
+  /**
+   * Bound a mutation push so it always settles.
+   *
+   * `SyncScheduler.syncUp` early-returns while `isSyncingUp` is true, and that
+   * flag only clears in the `finally` of the drain loop. A push whose RPC never
+   * settles (socket dropped mid-flight, response lost) therefore wedges the
+   * up-queue for the rest of the session: no retry, no error, no further
+   * mutation ever sent. A timeout turns that into an ordinary network failure,
+   * which `UpQueue.next` re-queues for the next trigger. The message deliberately
+   * contains "timed out" so `classifySyncError` treats it as `network` and
+   * retries rather than rolling the mutation back.
+   */
+  private withPushTimeout<T>(promise: Promise<T>, label: string): Promise<T> {
+    return withTimeout(
+      promise,
+      this.pushTimeoutMs,
+      `Mutation push timed out after ${this.pushTimeoutMs}ms (${label})`
+    );
+  }
+
   private async processUpEvent(event: UpEvent) {
     this.logger.debug(
-      { event, Category: 'spooky-client::SpookySync::processUpEvent' },
+      { event, Category: 'sp00ky-client::Sp00kySync::processUpEvent' },
       'Processing up event'
     );
-    console.log('xx1', event);
     switch (event.type) {
-      case 'create':
+      case 'create': {
         const dataKeys = Object.keys(event.data).map((key) => ({ key, variable: `data_${key}` }));
         const prefixedParams = Object.fromEntries(
           dataKeys.map(({ key, variable }) => [variable, event.data[key]])
         );
         const query = surql.seal(surql.createSet('id', dataKeys));
-        await this.remote.query(query, {
-          id: event.record_id,
-          ...prefixedParams,
-        });
+        await this.withPushTimeout(
+          this.remote.query(query, {
+            id: event.record_id,
+            ...prefixedParams,
+          }),
+          'create'
+        );
         break;
+      }
       case 'update':
-        await this.remote.query(`UPDATE $id MERGE $data`, {
-          id: event.record_id,
-          data: event.data,
-        });
+        await this.withPushTimeout(
+          this.remote.query(`UPDATE $id MERGE $data`, {
+            id: event.record_id,
+            data: event.data,
+          }),
+          'update'
+        );
         break;
       case 'delete':
-        await this.remote.query(`DELETE $id`, {
-          id: event.record_id,
-        });
+        await this.withPushTimeout(
+          this.remote.query(`DELETE $id`, {
+            id: event.record_id,
+          }),
+          'delete'
+        );
         break;
       default:
         this.logger.error(
-          { event, Category: 'spooky-client::SpookySync::processUpEvent' },
+          { event, Category: 'sp00ky-client::Sp00kySync::processUpEvent' },
           'processUpEvent unknown event type'
         );
         return;
     }
   }
 
+  /**
+   * A mutation the server accepted, reported once its outbox row is gone.
+   *
+   * Keeps the written row in the render set until its membership arrives.
+   * Without this the row is briefly in neither term of
+   * `(membership ∪ pendingWrites) − pendingDeletes` — the outbox delete is
+   * tied to the push, while membership waits on the SSP ingesting the row,
+   * materializing the view, writing the `_00_list_ref` edge and this client
+   * reading it back. The writer therefore watched its own comment appear,
+   * vanish, and return, while every other client showed it throughout.
+   */
+  private handleMutationSettled(event: UpEvent): void {
+    this.dataModule.noteWriteSettled(encodeRecordId(event.record_id), event.type);
+  }
+
   private async handleRollback(event: UpEvent, error: Error): Promise<void> {
     const recordId = encodeRecordId(event.record_id);
     const tableName =
-      event.type === 'create' && event.tableName
-        ? event.tableName
-        : extractTablePart(recordId);
+      event.type === 'create' && event.tableName ? event.tableName : extractTablePart(recordId);
 
     this.logger.warn(
       {
@@ -215,7 +1389,7 @@ export class SpookySync<S extends SchemaStructure> {
         recordId,
         tableName,
         error: error.message,
-        Category: 'spooky-client::SpookySync::handleRollback',
+        Category: 'sp00ky-client::Sp00kySync::handleRollback',
       },
       'Rolling back failed mutation'
     );
@@ -231,7 +1405,7 @@ export class SpookySync<S extends SchemaStructure> {
           this.logger.warn(
             {
               recordId,
-              Category: 'spooky-client::SpookySync::handleRollback',
+              Category: 'sp00ky-client::Sp00kySync::handleRollback',
             },
             'Cannot rollback update: no beforeRecord available. Down-sync will reconcile.'
           );
@@ -241,7 +1415,7 @@ export class SpookySync<S extends SchemaStructure> {
         this.logger.warn(
           {
             recordId,
-            Category: 'spooky-client::SpookySync::handleRollback',
+            Category: 'sp00ky-client::Sp00kySync::handleRollback',
           },
           'Delete rollback not implemented. Down-sync will reconcile.'
         );
@@ -253,11 +1427,26 @@ export class SpookySync<S extends SchemaStructure> {
       recordId,
       error: error.message,
     });
+
+    // Shared-tabs: the store rollback above already propagated to every tab
+    // via the ingest relay; additionally deliver the EVENT to the tab that
+    // owns the mutation so its UI (toasts, subscribeToRollbacks) fires there.
+    const mutationId = encodeRecordId(event.mutation_id);
+    const owner = mutationOwnerTabId(mutationId);
+    if (this.hub && owner && this.tabId && owner !== this.tabId) {
+      this.hub.sendTo(owner, {
+        type: 'mutation-rolled-back',
+        mutationId,
+        recordId,
+        eventType: event.type,
+        error: error.message,
+      });
+    }
   }
 
   private async processDownEvent(event: DownEvent) {
     this.logger.debug(
-      { event, Category: 'spooky-client::SpookySync::processDownEvent' },
+      { event, Category: 'sp00ky-client::Sp00kySync::processDownEvent' },
       'Processing down event'
     );
     switch (event.type) {
@@ -281,7 +1470,7 @@ export class SpookySync<S extends SchemaStructure> {
     const queryState = this.dataModule.getQueryByHash(hash);
     if (!queryState) {
       this.logger.warn(
-        { hash, Category: 'spooky-client::SpookySync::syncQuery' },
+        { hash, Category: 'sp00ky-client::Sp00kySync::syncQuery' },
         'Query not found'
       );
       return;
@@ -295,7 +1484,107 @@ export class SpookySync<S extends SchemaStructure> {
     if (!diff) {
       return;
     }
-    return this.syncEngine.syncRecords(diff);
+    return this.runSyncForQuery(hash, diff);
+  }
+
+  /**
+   * Run a sync for a single query while reflecting its fetch status. Marks the
+   * query `fetching` for the duration when the diff actually pulls records
+   * (added/updated), then resets to `idle` in a `finally` so a failed sync
+   * never leaves a query stuck `fetching`. Part A's notification coalescing
+   * means the single resulting UI update lands after this completes.
+   */
+  private async runSyncForQuery(hash: string, diff: RecordVersionDiff): Promise<void> {
+    // Don't let sync re-add a record the user just deleted locally. The remote
+    // delete is queued in the outbox, so until it's processed the server's
+    // `_00_list_ref` still lists the record — the diff then classifies it as
+    // `added` (present remotely, absent locally) and `syncRecords` re-fetches +
+    // re-inserts it, so a deleted database reappears a few seconds later. Drop
+    // any id with a pending local DELETE from the re-add paths. Once the remote
+    // delete lands, the pending row clears and the server drops it from
+    // `_00_list_ref`, so this guard naturally stops applying.
+    if (diff.added.length > 0 || diff.updated.length > 0) {
+      const pendingDeletes = await this.getPendingDeleteIds();
+      if (pendingDeletes.size > 0) {
+        diff = {
+          added: diff.added.filter((r) => !pendingDeletes.has(encodeRecordId(r.id))),
+          updated: diff.updated.filter((r) => !pendingDeletes.has(encodeRecordId(r.id))),
+          removed: diff.removed,
+        };
+      }
+    }
+
+    const fetching = diff.added.length + diff.updated.length > 0;
+    if (fetching) {
+      this.dataModule.beginFetching(hash);
+    }
+    try {
+      const { remoteFetchMs, stillRemoteIds } = await this.syncEngine.syncRecords(diff);
+      if (fetching) {
+        this.dataModule.recordRemoteFetch(hash, remoteFetchMs);
+      }
+      // Converge localArray to the authoritative remoteArray for ids that left
+      // the server's list_ref but still exist — a view-membership change, not a
+      // delete — so the poll's diff stops re-flagging them every tick (the `job:`
+      // churn). CRUCIAL: only converge after the id has been still-remote for
+      // several CONSECUTIVE rounds. A record that's merely mid-deletion is
+      // still-remote for ~one round (its delete hasn't committed when our
+      // existence check races it) and is gone the next round → it never reaches
+      // the threshold, so it's deleted normally instead of being stranded here.
+      if (stillRemoteIds.length > 0) {
+        const CONVERGE_AFTER = 3;
+        const toConverge: string[] = [];
+        for (const id of stillRemoteIds) {
+          const key = `${hash}:${id}`;
+          const n = (this.stillRemoteStreaks.get(key) ?? 0) + 1;
+          if (n >= CONVERGE_AFTER) {
+            this.stillRemoteStreaks.delete(key);
+            toConverge.push(id);
+          } else {
+            this.stillRemoteStreaks.set(key, n);
+          }
+        }
+        if (toConverge.length > 0) {
+          const qs = this.dataModule.getQueryByHash(hash);
+          const local = qs?.config.localArray;
+          if (local && local.length > 0) {
+            const drop = new Set(toConverge);
+            const next = local.filter(([id]) => !drop.has(id));
+            if (next.length !== local.length) {
+              await this.dataModule.updateQueryLocalArray(hash, next);
+            }
+          }
+        }
+      }
+    } finally {
+      if (fetching) {
+        // Land the coalesced result BEFORE flipping to idle: the final stream
+        // update sits on a debounce timer, and an `idle` that races ahead of it
+        // would let consumers treat a partially-filled window as authoritative.
+        try {
+          await this.dataModule.flushPendingStreamUpdate(hash);
+        } catch (err) {
+          this.logger.warn(
+            { err, hash, Category: 'sp00ky-client::Sp00kySync::runSyncForQuery' },
+            'Failed to flush pending stream update before idle'
+          );
+        }
+        this.dataModule.endFetching(hash);
+      }
+    }
+  }
+
+  /**
+   * Record ids with a pending local DELETE in the outbox (`_00_pending_mutations`).
+   * Sync must not re-fetch/re-insert these — the remote delete is async, so the
+   * server's `_00_list_ref` still lists them until it's processed, and the diff
+   * would otherwise resurrect a just-deleted record.
+   */
+  private async getPendingDeleteIds(): Promise<Set<string>> {
+    // Single implementation, shared with the render path: `materializeRecords`
+    // subtracts the same set so a row whose DELETE is still in the outbox is
+    // neither re-fetched here nor rendered there.
+    return (await this.dataModule.getPendingRecordIds()).deletes;
   }
 
   /**
@@ -303,26 +1592,46 @@ export class SpookySync<S extends SchemaStructure> {
    * @param mutations Array of UpEvents (create/update/delete) to enqueue.
    */
   public async enqueueMutation(mutations: UpEvent[]) {
+    // Follower: the outbox rows are already committed in the SHARED store (the
+    // mutation tx went through the leader's worker); only the leader drains,
+    // so hand over the ids instead of queueing locally. A notify lost in a
+    // failover window is covered by the new leader's loadFromDatabase.
+    if (this.tabRole === 'follower') {
+      for (const m of mutations) {
+        this.forwarder?.mutationEnqueued(encodeRecordId(m.mutation_id));
+      }
+      return;
+    }
     this.scheduler.enqueueMutation(mutations);
   }
 
   private async registerQuery(queryHash: string) {
+    // Hold `fetching` across the WHOLE registration (remote view creation +
+    // initial sync + post-sync notify). A query is born `fetching` in
+    // createNewQuery; this refcounted cycle is what resolves it to `idle` — so
+    // consumers (e.g. useQuery's `isSettled`) never see an idle query whose
+    // window is still empty/partially materialized.
+    this.dataModule.beginFetching(queryHash);
     try {
       this.logger.debug(
-        { queryHash, Category: 'spooky-client::SpookySync::registerQuery' },
+        { queryHash, Category: 'sp00ky-client::Sp00kySync::registerQuery' },
         'Register Query state'
       );
       await this.createRemoteQuery(queryHash);
       await this.syncQuery(queryHash);
-      // Always notify after sync completes — handles empty result sets
-      // where no stream updates fire but the UI needs to stop loading
+      // Land any still-debounced stream result, then always notify — handles
+      // empty result sets where no stream updates fire but the UI needs to
+      // stop loading.
+      await this.dataModule.flushPendingStreamUpdate(queryHash);
       await this.dataModule.notifyQuerySynced(queryHash);
     } catch (e) {
       this.logger.error(
-        { err: e, Category: 'spooky-client::SpookySync::registerQuery' },
+        { err: e, Category: 'sp00ky-client::Sp00kySync::registerQuery' },
         'registerQuery error'
       );
       throw e;
+    } finally {
+      this.dataModule.endFetching(queryHash);
     }
   }
 
@@ -331,15 +1640,15 @@ export class SpookySync<S extends SchemaStructure> {
 
     if (!queryState) {
       this.logger.warn(
-        { queryHash, Category: 'spooky-client::SpookySync::createRemoteQuery' },
+        { queryHash, Category: 'sp00ky-client::Sp00kySync::createRemoteQuery' },
         'Query to register not found'
       );
       throw new Error('Query to register not found');
     }
-    // Delegate to remote function which handles DBSP registration & persistence
+    // Delegate to remote function which handles DBSP registration & persistence.
+    // clientId is set server-side from session::id() — see fn::query::register.
     await this.remote.query('fn::query::register($config)', {
       config: {
-        clientId: this.clientId,
         id: queryState.config.id,
         surql: queryState.config.surql,
         params: queryState.config.params,
@@ -347,18 +1656,26 @@ export class SpookySync<S extends SchemaStructure> {
       },
     });
 
-    const [items] = await this.remote.query<[{ out: RecordId<string>; version: number }[]]>(
-      surql.selectByFieldsAnd('_spooky_list_ref', ['in'], ['out', 'version']),
-      {
-        in: queryState.config.id,
-      }
-    );
+    // Initial materialized-view fetch — pull from the same per-user
+    // `_00_list_ref_user_<id>` (or global `_00_list_ref` in single
+    // mode) that the LIVE subscription listens on, so the two stay in
+    // sync. `parent IS NONE` excludes subquery entries; the
+    // `localArray` cache only tracks primary records.
+    const listRefTbl = this.listRefTable();
+    // `rowCount` rides along: it is written by the SSP in the same statement
+    // that registers the view, BEFORE the edges are flushed, so it is the only
+    // way to tell "this query is empty" from "its edges have not landed yet".
+    const [items, serverRowCount] = await this.remote.query<
+      [{ out: RecordId<string>; version: number }[], number | null]
+    >(`${buildListRefSelect(listRefTbl)};\n${buildQueryRowCountSelect()}`, {
+      in: queryState.config.id,
+    });
 
     this.logger.trace(
       {
         queryId: encodeRecordId(queryState.config.id),
         items,
-        Category: 'spooky-client::SpookySync::createRemoteQuery',
+        Category: 'sp00ky-client::Sp00kySync::createRemoteQuery',
       },
       'Got query record version array from remote'
     );
@@ -369,42 +1686,184 @@ export class SpookySync<S extends SchemaStructure> {
       {
         queryId: encodeRecordId(queryState.config.id),
         array,
-        Category: 'spooky-client::SpookySync::createRemoteQuery',
+        Category: 'sp00ky-client::Sp00kySync::createRemoteQuery',
       },
       'createdRemoteQuery'
     );
 
     if (array) {
       /// Incantation existed already
-      await this.dataModule.updateQueryRemoteArray(queryHash, array);
+      await this.dataModule.updateQueryRemoteArray(queryHash, array, { serverRowCount });
+    }
+
+    // Pull the bodies of any `.related()` subquery children into the local
+    // cache. The primary fetch above (`parent IS NONE`) tracks only window
+    // rows, so without this a cold-reload re-materialization of the
+    // correlated surql finds no child rows and related fields come back
+    // empty. Best-effort: never fail registration over it.
+    await this.syncSubqueryChildren(queryHash).catch((err) => {
+      this.logger.info(
+        {
+          err: (err as Error)?.message ?? err,
+          queryHash,
+          Category: 'sp00ky-client::Sp00kySync::createRemoteQuery',
+        },
+        'Subquery child sync failed during registration; poll will retry'
+      );
+    });
+  }
+
+  /**
+   * Sync the BODIES of a `.related()` query's subquery child rows into the
+   * local cache, separately from the primary window array. The SSP writes
+   * each matched child as a `_00_list_ref` edge tagged `parent`/`parent_rel`;
+   * `buildSubqueryListRefSelect` pulls those `out`+`version` pairs (any
+   * nesting depth). We diff against the in-memory `subqueryRemoteArray` and
+   * fetch added/updated bodies through the SyncEngine — which `saveBatch`s
+   * them into the local DB AND the in-browser SSP, whose subquery-table
+   * dependency then re-materializes the parent view (no explicit notify).
+   *
+   * Deletion safety: we pass `removed: []` deliberately. A child body can be
+   * shared by other queries; letting `handleRemovedRecords` delete one that
+   * merely left THIS query's child set would clobber data another query still
+   * shows. Genuine record deletes flow through the normal delete path; a
+   * lingering orphan body is invisible (the correlated WHERE stops matching).
+   *
+   * Kept off `runSyncForQuery` on purpose so child fetches never flip the
+   * query to `fetching` or skew its DevTools timings.
+   */
+  private async syncSubqueryChildren(queryHash: string): Promise<void> {
+    const queryState = this.dataModule.getQueryByHash(queryHash);
+    if (!queryState) return;
+
+    const listRefTbl = this.listRefTable();
+    const [items] = await this.remote.query<[{ out: RecordId<string>; version: number }[]]>(
+      buildSubqueryListRefSelect(listRefTbl),
+      { in: queryState.config.id }
+    );
+    if (!Array.isArray(items)) return;
+
+    const fresh: RecordVersionArray = items.map((item) => [encodeRecordId(item.out), item.version]);
+    const prev = queryState.config.subqueryRemoteArray ?? [];
+    if (recordVersionArraysEqual(fresh, prev)) return; // idempotent: nothing new
+
+    const diff = diffRecordVersionArray(prev, fresh);
+    if (diff.added.length > 0 || diff.updated.length > 0) {
+      await this.syncEngine.syncRecords({
+        added: diff.added,
+        updated: diff.updated,
+        removed: [], // never delete child bodies here — see method doc
+      });
     }
+    // In-memory only — child rows must never enter the persisted primary array.
+    queryState.config.subqueryRemoteArray = fresh;
   }
 
-  private async heartbeatQuery(queryHash: string) {
+  public async heartbeatQuery(queryHash: string) {
     const queryState = this.dataModule.getQueryByHash(queryHash);
     if (!queryState) {
       this.logger.warn(
-        { queryHash, Category: 'spooky-client::SpookySync::heartbeatQuery' },
+        { queryHash, Category: 'sp00ky-client::Sp00kySync::heartbeatQuery' },
         'Query to register not found'
       );
       throw new Error('Query to register not found');
     }
-    await this.remote.query('fn::query::heartbeat($id)', {
+    // `fn::query::heartbeat` is an `UPDATE $id SET ...`. On a record that no
+    // longer exists that matches nothing and returns an empty array — it does
+    // NOT recreate the row. So an unchecked heartbeat is indistinguishable from
+    // a successful one, and a client whose row was reclaimed keeps beating
+    // against nothing forever: no membership, no edges, no re-registration.
+    // The page renders as if the data were deleted ("Game not found").
+    //
+    // A live query's row is reclaimed more easily than it looks. The sweep
+    // expires on `lastActiveAt + ttl`, and this heartbeat runs on a timer that
+    // browsers throttle hard in background tabs — so a second window left idle
+    // past its TTL is the ordinary way to get here, not an edge case. Until
+    // canary.194 the sweep could not actually remove the in-memory view (it
+    // looked it up under the other of the two query-id spellings), which masked
+    // this: the view survived its own row. Now reclamation is real, so the
+    // client has to notice and rebuild.
+    const result = await this.remote.query('fn::query::heartbeat($id)', {
       id: queryState.config.id,
     });
+    const updated = Array.isArray(result) ? result[0] : undefined;
+    const rowGone = Array.isArray(updated) && updated.length === 0;
+    if (!rowGone) return;
+
+    this.logger.warn(
+      {
+        queryHash,
+        id: String(queryState.config.id),
+        Category: 'sp00ky-client::Sp00kySync::heartbeatQuery',
+      },
+      'Query row was reclaimed while still in use; re-registering'
+    );
+    // Re-register rather than recreate the row here: the row alone is useless
+    // without the SSP view behind it, and only registration rebuilds the view,
+    // republishes `_00_list_ref` and writes `rowCount`.
+    this.enqueueDownEvent({ type: 'register', payload: { hash: queryHash } });
   }
 
+  // Eager teardown of a deregistered query's remote `_00_query` view (opt-in,
+  // e.g. a viewport-windowed list cancelling an off-screen window). Query ids
+  // are a deterministic hash of (surql+params), so a release racing a
+  // scroll-back re-register (same id) could nuke a freshly-recreated view —
+  // hence two guards: abort if a subscriber reappeared BEFORE the release;
+  // re-register if one reappears DURING the release's network await. Tolerant
+  // of a missing/already-gone query (no throw).
+  //
+  // Releases via `fn::query::unsubscribe` rather than deleting the row
+  // outright. A `_00_query` row can be shared by several sessions of the same
+  // user, so a bare `DELETE` would tear the view — and every `_00_list_ref`
+  // edge hanging off it — out from under other live tabs. The function drops
+  // only this session from `subscribers` and deletes the row when it was the
+  // last one. (The old `DELETE $id` was harmless in practice only because the
+  // table granted no delete permission and it silently affected zero rows.)
   private async cleanupQuery(queryHash: string) {
     const queryState = this.dataModule.getQueryByHash(queryHash);
-    if (!queryState) {
-      this.logger.warn(
-        { queryHash, Category: 'spooky-client::SpookySync::cleanupQuery' },
-        'Query to register not found'
-      );
-      throw new Error('Query to register not found');
+    if (!queryState) return; // already torn down / never registered
+
+    // Re-subscribed before the queued cleanup ran → keep everything as-is.
+    if (this.dataModule.hasSubscribers(queryHash)) return;
+
+    // EAGER REMOTE RELEASE IS DISABLED. Deliberate, and not a leak: the TTL
+    // sweep reclaims the row and its edges on `lastActiveAt + ttl`, which is the
+    // ONLY reclamation that has ever actually run in production.
+    //
+    // Until canary.190 `_00_query` granted no delete permission, so the bare
+    // `DELETE $id` this used to issue affected zero rows. .190 granted delete
+    // and .191 wired `fn::query::unsubscribe`, which made teardown real for the
+    // first time -- and the guards above are best-effort by construction
+    // (`hasSubscribers` can be momentarily false during a rebind or a windowed
+    // list re-flow). Every misfire that had been silently inert for months
+    // became a live delete of the row AND every `_00_list_ref` edge on it.
+    //
+    // That matches a report of chat suddenly rendering raw record ids instead
+    // of users, with the message list re-flowing underneath. Server state was
+    // measured intact at the time (`rowCount` equalled the actual edge count on
+    // every row), so the damage is on the client side of a teardown, not in the
+    // materialization.
+    //
+    // Re-enable only together with a repair path that can re-fetch subquery
+    // child bodies whose `subqueryRemoteArray` entry claims they are already
+    // synced -- otherwise a torn-down-and-recreated view never restores the
+    // related records it dropped, because the idempotence check skips them.
+    if (this.releaseQueriesEagerly) {
+      await this.remote.query('fn::query::unsubscribe($id)', {
+        id: queryState.config.id,
+      });
+
+      // Re-subscribed while we awaited the release → re-register. Covers both
+      // outcomes: if we were the last subscriber the remote view is gone and
+      // this recreates it, and if it survived for other sessions this re-adds
+      // us to `subscribers` so our heartbeats keep counting.
+      if (this.dataModule.hasSubscribers(queryHash)) {
+        this.enqueueDownEvent({ type: 'register', payload: { hash: queryHash } });
+        return;
+      }
     }
-    await this.remote.query(`DELETE $id`, {
-      id: queryState.config.id,
-    });
+
+    // No subscribers throughout → safe to free the local view + state.
+    this.dataModule.finalizeDeregister(queryHash);
   }
 }