@spooky-sync/core 0.0.1-canary.32 → 0.0.1-canary.321

package/src/services/database/sqlite-cache-engine.ts

@@ -0,0 +1,1372 @@
+import { applyPatch, type Operation } from 'fast-json-patch';
+import { DEFAULT_LOCAL_OP_TIMEOUT_MS, LocalOpTimeoutError } from './errors';
+import { withTimeout } from '../../utils/index';
+import type { QueryPlan, RelationPlan, WhereNode } from '@spooky-sync/query-builder';
+import {
+  renderOrderSql,
+  renderWhereSql,
+  reviveRow,
+  serializeMergeRow,
+  serializeRow,
+  project,
+  projectedDataSql,
+} from './sqlite-plan-sql';
+import type { Logger } from '../logger/index';
+import type { Sp00kyConfig, StorageHealth } from '../../types';
+import type { SnapshotMeta, StoredSnapshot } from './cache-engine';
+import type { SealedQuery } from '../../utils/surql';
+import { resolveRelations, stableKey } from './relation-resolver';
+import {
+  createDatabaseEventSystem,
+  DatabaseEventTypes,
+  type DatabaseEventSystem,
+} from './events/index';
+import { StaleEpochError } from './local';
+import { translateSurql, tableOf, setPath, getPath, type SqlOp } from './surql-translate';
+import {
+  BrokerPortClosedError,
+  PortSqliteTransport,
+  WorkerSqliteTransport,
+  type SqliteTransport,
+} from './sqlite-transport';
+import { PROMOTION_OPEN_OPTIONS } from './sqlite-open';
+import type { EngineTx, Id, LocalStore, OrderBy, RelationFetch, Row } from './cache-engine';
+import type { EngineStorageDiagnostics } from '../../modules/devtools/storage-info';
+
+/**
+ * The statement result a pure-write op contributes to a query's results array.
+ * Single source of truth shared by the per-op path (`execOp`) and the batched
+ * fast path in `query()`, so the two can never diverge: a caller that reads a
+ * statement's output sees the same shape whether or not the transaction took the
+ * batch fast path. In particular `create()` compiles to an all-upsert tx
+ * (`createSet` + `createMutation`) and reads `resultIndex:0` for the new row and
+ * its id — the fast path previously returned empty arrays there, so the row (and
+ * its id) was lost and the reconcile crashed in `encodeRecordId`.
+ *
+ * An upsert echoes the written row (`{...data, id}`) with no read-back — the
+ * full merged row is only materialized for a LET-wrapped upsert (see the 'let'
+ * case). delete/deleteAll yield `[]`; noop yields `null`.
+ */
+/**
+ * The `_00_*` internal tables the client relies on. The LocalMigrator DEFINEs
+ * them, but every DEFINE lowers to a noop on the SQLite engine, so they must be
+ * created physically at open (see `openInternal`) or a read-before-first-write
+ * on a fresh bucket throws "no such table". Keep in sync with the systemSchema
+ * block in `local-migrator.ts`.
+ */
+const SNAPSHOT_TABLE = '_00_circuit_snapshot';
+
+const SYSTEM_TABLES = [
+  '_00_stream_processor_state',
+  // In-browser circuit snapshot: one BLOB row (`circuit`) + a JSON meta row.
+  // Excluded from DevTools' table listing, since its data column is not JSON.
+  '_00_circuit_snapshot',
+  '_00_view',
+  '_00_window',
+  '_00_failed_mutations',
+  '_00_schema',
+  '_00_pending_mutations',
+  // Debounced patches still inside their delay; read on every boot.
+  '_00_pending_writes',
+  // Blob cache manifest. Read before its first write on every boot (reconcile
+  // asks for the ids it found in OPFS), so it has to exist up front like the
+  // rest — otherwise the very first reconcile throws and the cache runs cold.
+  '_00_blob',
+  // Server-written, synced-down meta tables (see meta_tables_client.surql).
+  // DEFINE is a noop on this engine, so without seeding them here their synced
+  // rows have no local table to land in: feature flags silently fall back to
+  // defaults and app-release update notifications never show.
+  '_00_user_feature',
+  '_00_app_release',
+] as const;
+
+export function pureWriteOpResult(op: SqlOp): unknown {
+  switch (op.kind) {
+    case 'upsert':
+      return { ...op.data, id: stableKey(op.id) };
+    case 'noop':
+      return null;
+    default:
+      return [];
+  }
+}
+
+/**
+ * Local cache backend on official SQLite-WASM in a dedicated Worker (see
+ * `sqlite-worker.ts`), with OPFS SAHPool persistence. Storage model: one table
+ * per schema table, `id TEXT PRIMARY KEY, data TEXT` where `data` is the row as
+ * JSON. Filtering/ordering use `json_extract`. Relations are decomposed by the
+ * shared {@link resolveRelations} — identical to the SurrealDB backend.
+ *
+ * Value normalization (JSON round-trip):
+ * - `Uint8Array`/bytes → `{ "__u8": <base64> }` (CRDT snapshots survive).
+ * - Record links / ids → their `table:id` string form (so `json_extract`
+ *   comparisons and `IN` matching are consistent). NOTE: link fields therefore
+ *   read back as strings, not `RecordId` instances — the one shape difference
+ *   from the SurrealDB backend, to be closed with schema-driven revival + an
+ *   oracle E2E in the browser.
+ */
+export class SqliteCacheEngine implements LocalStore {
+  /** The wire to the SQLite worker: an owned dedicated Worker (solo/leader) or
+   *  a MessagePort into another tab's worker (follower). See sqlite-transport. */
+  private transport: SqliteTransport | null = null;
+  private storeEpoch = 0;
+  private knownTables = new Set<string>();
+  private useOpfs: boolean;
+  /** Whether `select` runs as one worker round-trip (plan executed in-worker).
+   *  Flipped off at runtime if the worker script predates the `select` op
+   *  (stale cached bundle) — degrade to the legacy multi-hop path, don't break. */
+  private workerSelect: boolean;
+  /** What `workerSelect` was at construction, so DevTools can tell a runtime
+   *  downgrade (configured true, effective false) from a configured-off. */
+  private workerSelectConfigured: boolean;
+  private events: DatabaseEventSystem = createDatabaseEventSystem();
+  private bucketId = 'anon';
+  /** Deadline for one worker round trip; see `localOpTimeoutMs`. */
+  private readonly localOpTimeoutMs: number;
+  /** Durability of the local store, set on every open. A plain Set of callbacks
+   *  rather than a `DatabaseEventSystem` event: this changes at most once per
+   *  open, and the typed event map is about query traffic. */
+  private storageHealthValue: StorageHealth = { status: 'unknown', fallback: false };
+  private storageHealthSubs = new Set<(health: StorageHealth) => void>();
+  /** Schemaless — tables are created lazily on first write; no migrator. */
+  readonly usesSurqlSchema = false;
+
+  readonly engineKind = 'sqlite' as const;
+
+  /** Shared-tabs mode: the engine's transport is swapped at runtime by the
+   *  TabsCoordinator (owner worker as leader, leader's port as follower). */
+  private shared: boolean;
+  /** Leaderless parking (shared mode): ops entering the opQueue await this
+   *  gate until a new role lands or the timeout rejects them. */
+  private roleGate: { promise: Promise<void>; release: () => void } | null = null;
+  private roleGateTimer: ReturnType<typeof setTimeout> | null = null;
+
+  constructor(
+    private config: Sp00kyConfig<any>['database'],
+    private logger: Logger,
+    opts: { useOpfs?: boolean; workerSelect?: boolean; shared?: boolean } = {}
+  ) {
+    this.localOpTimeoutMs = Math.max(0, config.localOpTimeoutMs ?? DEFAULT_LOCAL_OP_TIMEOUT_MS);
+    this.useOpfs = opts.useOpfs ?? true;
+    this.workerSelect = opts.workerSelect ?? config.workerSelect ?? true;
+    this.workerSelectConfigured = this.workerSelect;
+    this.shared = opts.shared ?? false;
+  }
+
+  get epoch(): number {
+    return this.storeEpoch;
+  }
+
+  get currentBucketId(): string {
+    return this.bucketId;
+  }
+
+  get storageHealth(): StorageHealth {
+    return this.storageHealthValue;
+  }
+
+  /** Fires immediately with the current snapshot (the store opens during
+   *  `connect()`, before app components mount, so a late subscriber must still
+   *  learn a fallback happened), then on every change. */
+  subscribeToStorageHealth(cb: (health: StorageHealth) => void): () => void {
+    cb(this.storageHealthValue);
+    this.storageHealthSubs.add(cb);
+    return () => {
+      this.storageHealthSubs.delete(cb);
+    };
+  }
+
+  private setStorageHealth(health: StorageHealth): void {
+    this.storageHealthValue = health;
+    for (const cb of this.storageHealthSubs) cb(health);
+  }
+
+  getConfig(): Sp00kyConfig<any>['database'] {
+    return this.config;
+  }
+
+  /**
+   * Storage numbers for the DevTools Storage tab. Uses {@link call} so the
+   * reads serialize with regular traffic (no SQLITE_BUSY). Never throws — the
+   * worker may be mid bucket-switch; a failure lands in `error` instead.
+   */
+  async getStorageDiagnostics(opts?: { tableCounts?: boolean }): Promise<EngineStorageDiagnostics> {
+    const diag: EngineStorageDiagnostics = {
+      engine: 'sqlite',
+      bucketId: this.bucketId,
+      useOpfs: this.useOpfs,
+      workerSelectConfigured: this.workerSelectConfigured,
+      workerSelectEffective: this.workerSelect,
+    };
+    try {
+      const { rows } = await this.call<{ rows: { bytes: number; freelist: number }[] }>('exec', {
+        sql:
+          'SELECT (SELECT * FROM pragma_page_count()) * (SELECT * FROM pragma_page_size()) AS bytes, ' +
+          '(SELECT * FROM pragma_freelist_count()) * (SELECT * FROM pragma_page_size()) AS freelist',
+      });
+      diag.dbSizeBytes = rows?.[0]?.bytes;
+      diag.freelistBytes = rows?.[0]?.freelist;
+      if (opts?.tableCounts) {
+        const { rows: tables } = await this.call<{ rows: { name: string }[] }>('exec', {
+          sql: "SELECT name FROM sqlite_master WHERE type='table' ORDER BY name",
+        });
+        const names = (tables ?? []).map((r) => r.name).filter((n) => n !== SNAPSHOT_TABLE);
+        if (names.length) {
+          // Names come from sqlite_master itself; double-quoting is enough.
+          const sql = names
+            .map((n) => `SELECT '${n.replace(/'/g, "''")}' AS t, COUNT(*) AS n FROM "${n.replace(/"/g, '""')}"`)
+            .join(' UNION ALL ');
+          const { rows: counts } = await this.call<{ rows: { t: string; n: number }[] }>('exec', {
+            sql,
+          });
+          diag.tableCounts = (counts ?? []).map((r) => ({ table: r.t, rows: r.n }));
+        } else {
+          diag.tableCounts = [];
+        }
+      }
+    } catch (e) {
+      diag.error = e instanceof Error ? e.message : String(e);
+    }
+    return diag;
+  }
+
+  getEvents(): DatabaseEventSystem {
+    return this.events;
+  }
+
+  getClient(): unknown {
+    throw new Error('SqliteCacheEngine has no SurrealDB client (getClient is unavailable).');
+  }
+
+  /**
+   * See `LocalStore.dropBucket`. Each bucket is its own OPFS SAH pool, whose
+   * directory sqlite-wasm names `.<pool name>` (`sp00ky-<bucket>`, see
+   * sqlite-open.ts). Fails harmlessly while a worker still holds its handles.
+   */
+  async dropBucket(bucketId: string): Promise<void> {
+    if (bucketId === this.bucketId) throw new Error('Cannot drop the open bucket.');
+    const storage = typeof navigator !== 'undefined' ? navigator.storage : undefined;
+    if (!storage?.getDirectory) return;
+    const root = await storage.getDirectory();
+    try {
+      await root.removeEntry(`.sp00ky-${bucketId}`, { recursive: true });
+    } catch (error) {
+      if ((error as DOMException)?.name !== 'NotFoundError') throw error;
+    }
+  }
+
+  /** LocalStore alias; SQLite has no in-flight gate, so this maps to a rebuild. */
+  switchStore(bucketId: string): Promise<void> {
+    return this.switchBucket(bucketId);
+  }
+
+  /** SQLite has no switch gate window; the epoch bump alone fences stale writes. */
+  beginSwitch(): () => void {
+    return () => {};
+  }
+
+  // ---- worker plumbing -----------------------------------------------------
+
+  /** Serializes every worker op so reads/writes never overlap at the VFS layer
+   *  (overlapping ops trip SQLITE_BUSY). Mirrors the SurrealDB engine's
+   *  single-flight query queue. (The worker keeps its own chain too, for
+   *  multi-client mode; this one additionally provides the queue-wait stat and
+   *  the boot/switch atomicity below.) */
+  private opQueue: Promise<unknown> = Promise.resolve();
+
+  private call<T = any>(type: string, payload?: unknown): Promise<T> {
+    const enqueuedAt = performance.now();
+    const run = async () => {
+      // Leaderless window (shared mode): park behind the role gate instead of
+      // failing; a new leader releases it, the timeout rejects it.
+      if (this.roleGate) await this.roleGate.promise;
+      // Time spent waiting behind other ops in the queue, not doing work.
+      getStats().queueWaitMs += performance.now() - enqueuedAt;
+      return this.rawCall<T>(type, payload);
+    };
+    const result = this.opQueue.then(run, run);
+    // Keep the chain alive regardless of individual failures.
+    this.opQueue = result.then(
+      () => undefined,
+      () => undefined
+    );
+    return result;
+  }
+
+  /**
+   * Ops dispatched to the worker and not yet answered. Role transitions run on
+   * their own chain (see {@link transitionChain}), so they can start while the
+   * opQueue still has an op at the worker; tearing the transport down under it
+   * would reject that op — and its caller may be a query whose only fetch this
+   * was. {@link drainInFlight} lets a deliberate teardown wait them out.
+   */
+  private inFlightCalls = new Set<Promise<unknown>>();
+
+  /** Wait for dispatched ops to answer before a deliberate transport teardown.
+   *  Bounded: a wedged worker must not block the role change forever. */
+  private async drainInFlight(timeoutMs = 2_000): Promise<void> {
+    if (this.inFlightCalls.size === 0) return;
+    const settled = Promise.all([...this.inFlightCalls].map((p) => p.catch(() => undefined)));
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    try {
+      await Promise.race([
+        settled,
+        new Promise<void>((resolve) => {
+          timer = setTimeout(resolve, timeoutMs);
+        }),
+      ]);
+    } finally {
+      if (timer) clearTimeout(timer);
+    }
+  }
+
+  private rawCall<T = any>(type: string, payload?: unknown): Promise<T> {
+    if (!this.transport) throw new Error('SqliteCacheEngine: not connected');
+    // --- instrumentation: live, inspectable via `globalThis.__sqliteStats` ---
+    const s = getStats();
+    s.roundTrips++;
+    s.byType[type] = (s.byType[type] ?? 0) + 1;
+    if (type === 'batch' && Array.isArray(payload)) {
+      s.batchStatements += payload.length;
+      s.maxBatch = Math.max(s.maxBatch, payload.length);
+    }
+    s.inFlight++;
+    s.maxInFlight = Math.max(s.maxInFlight, s.inFlight);
+    if (this.transport.kind === 'port') s.proxiedOps = (s.proxiedOps ?? 0) + 1;
+    const sentAt = performance.now();
+    // Track the dispatch, not the returned promise: attaching a handler to
+    // `result` would mark it handled and swallow the unhandled-rejection
+    // reports that surface a caller which forgot to catch.
+    let settle!: () => void;
+    const tracked = new Promise<void>((res) => {
+      settle = res;
+    });
+    this.inFlightCalls.add(tracked);
+    const finish = () => {
+      this.inFlightCalls.delete(tracked);
+      settle();
+    };
+    // Deadline per op. The transport parks a call until the worker replies,
+    // and a worker starved behind a long op (or wedged on an unbounded lock
+    // check) answered never; every caller up the stack then waited forever.
+    // Reject the CALLER only: the op is still queued in the worker and its
+    // late reply is dropped by the transport (no pending entry left to
+    // resolve). Deliberately no teardown here - a slow worker is not a dead
+    // one, and killing it mid-write would cost more than the wait.
+    const timeoutMs = this.localOpTimeoutMs;
+    const result = withTimeout(
+      this.transport.call<T>(type, payload),
+      timeoutMs,
+      () => new LocalOpTimeoutError(type, timeoutMs)
+    ).then(
+      (v: T) => {
+        finish();
+        s.inFlight--;
+        // Split the round-trip: `wt` is time inside the worker's handler,
+        // the remainder is postMessage + scheduling overhead.
+        const wt = (v as { wt?: unknown } | null)?.wt;
+        if (typeof wt === 'number') {
+          s.workerMs += wt;
+          s.rpcOverheadMs += Math.max(0, performance.now() - sentAt - wt);
+        }
+        return v;
+      },
+      (e: unknown) => {
+        finish();
+        s.inFlight--;
+        if (e instanceof LocalOpTimeoutError) {
+          s.timeouts = (s.timeouts ?? 0) + 1;
+          this.logger.error(
+            { type, timeoutMs, Category: 'sp00ky-client::SqliteCacheEngine::rawCall' },
+            'Local store op did not answer within its deadline; rejecting the caller'
+          );
+        }
+        throw e;
+      }
+    );
+    return result;
+  }
+
+  /**
+   * Spawn the worker and open `bucketId`'s DB. Uses {@link rawCall} (NOT
+   * {@link call}) so it can run as the body of an already-queued opQueue entry
+   * without re-queuing onto itself. Callers must run it through the opQueue.
+   */
+  /** Seam for tests: swap in a fake transport instead of a real Worker. */
+  protected createTransport(): SqliteTransport {
+    return new WorkerSqliteTransport(this.logger);
+  }
+
+  private async openInternal(
+    bucketId: string,
+    extras?: {
+      workerLockName?: string;
+      openOptions?: { maxAttempts?: number; backoffMs?: number[]; disallowMemoryFallback?: boolean };
+    }
+  ): Promise<void> {
+    if (!this.transport || this.transport.kind !== 'worker' || !this.transport.connected) {
+      this.transport = this.createTransport();
+      if (this.transport instanceof WorkerSqliteTransport) {
+        // The worker fencing itself (leadership stolen from a frozen tab that
+        // thawed) is a leader-loss: park ops until the broker re-adopts us.
+        this.transport.onLockLost = (reason) => {
+          this.transport = null;
+          this.storeEpoch++;
+          this.openRoleGate(`worker fenced: ${reason}`);
+        };
+      }
+    }
+    // Seed the `_00_*` system tables as part of `open` (worker-side, one round
+    // trip). The LocalMigrator DEFINEs them, but `translateSurql` lowers every
+    // DEFINE to a noop on this engine (SQLite has no DDL vocabulary), so
+    // provisioning never actually creates them — they were only made lazily on
+    // first WRITE. A fresh bucket (e.g. right after signup) that READS one first
+    // (the sync layer selects `_00_query` before any row lands) hit
+    // "no such table: _00_query" and the client wedged on "Loading database".
+    // Creating them inside `open` guarantees any access order is safe without
+    // adding ops to the engine's queue.
+    // `opfsError` is absent from a worker bundle older than this field, which
+    // just reads as "no reason given" rather than breaking the open.
+    const { persisted, opfsError } = await this.rawCall<{
+      persisted: boolean;
+      opfsError?: string;
+    }>('open', {
+      dbName: bucketId,
+      useOpfs: this.useOpfs,
+      systemTables: SYSTEM_TABLES,
+      ...(extras?.workerLockName ? { workerLockName: extras.workerLockName } : {}),
+      ...(extras?.openOptions ? { openOptions: extras.openOptions } : {}),
+    });
+    this.knownTables.clear();
+    for (const t of SYSTEM_TABLES) this.knownTables.add(t);
+    this.bucketId = bucketId;
+    // Durability was requested but could not be had: the store is in RAM, so it
+    // loses local writes on reload and can OOM a wasm-heavy renderer. Report it
+    // (the worker also console.errors, since host apps may run pino at `fatal`)
+    // and publish it so the app can warn the user.
+    const fellBack = this.useOpfs && !persisted;
+    // Omit `error` rather than setting it to `undefined`: the devtools
+    // serializer renders an undefined value as the STRING 'undefined'.
+    const health: StorageHealth = {
+      status: persisted ? 'persistent' : 'memory',
+      fallback: fellBack,
+    };
+    if (fellBack && opfsError) health.error = opfsError;
+    if (this.shared) health.role = this.roleLabel;
+    this.setStorageHealth(health);
+    const stats = getStats();
+    stats.persisted = persisted;
+    if (fellBack && opfsError) stats.opfsError = opfsError;
+    else delete stats.opfsError;
+    if (fellBack) {
+      this.logger.error(
+        { bucketId, opfsError, Category: 'sp00ky-client::SqliteCacheEngine::connect' },
+        'SQLite OPFS persistence failed; store is IN MEMORY and will not survive reload'
+      );
+    } else {
+      this.logger.info(
+        { bucketId, persisted, Category: 'sp00ky-client::SqliteCacheEngine::connect' },
+        persisted ? 'SQLite OPFS store opened' : 'SQLite in-memory store opened (as configured)'
+      );
+    }
+  }
+
+  /** Enqueue `fn` as a single serialized opQueue entry (mirrors {@link call}'s
+   *  chaining) so it can't interleave with reads/writes at the worker. */
+  private enqueue<T>(fn: () => Promise<T>): Promise<T> {
+    const result = this.opQueue.then(fn, fn);
+    this.opQueue = result.then(
+      () => undefined,
+      () => undefined
+    );
+    return result;
+  }
+
+  async connect(bucketId: string): Promise<void> {
+    // Serialize through the opQueue so an op racing boot can't dispatch to a
+    // half-open worker.
+    await this.enqueue(() => this.openInternal(bucketId));
+  }
+
+  async switchBucket(bucketId: string): Promise<void> {
+    this.storeEpoch++;
+    // Run close → terminate → reopen as ONE opQueue entry. Previously `close`
+    // and the new `open` were separate entries, so a query's `exec` (e.g. the
+    // query re-registration fired by an auth/bucket change) could slot in
+    // between and dispatch to the just-closed worker → "sqlite: DB not open".
+    // As a single entry, every other op runs fully before the close or after
+    // the reopen — never against a closed DB.
+    await this.enqueue(async () => {
+      if (this.transport) {
+        try {
+          await this.rawCall('close');
+        } catch {
+          /* ignore */
+        }
+        this.transport.close('bucket switch');
+        this.transport = null;
+      }
+      await this.openInternal(bucketId);
+    });
+  }
+
+  async close(): Promise<void> {
+    if (!this.transport) return;
+    try {
+      await this.call('close');
+    } catch {
+      /* ignore */
+    }
+    this.transport.close('engine closed');
+    this.transport = null;
+  }
+
+  // ---- shared-tabs role modes ------------------------------------------------
+  // The TabsCoordinator drives these; solo mode never touches them. The engine
+  // object is never replaced across role changes, so `storeEpoch` stays one
+  // monotonic per-tab counter and every existing fencing consumer keeps
+  // working unchanged.
+
+  /** Which role the current transport represents, for StorageHealth. */
+  private roleLabel: 'leader' | 'follower' | 'solo' = 'solo';
+  /** True once this engine has had a usable store at least once; role changes
+   *  after that point invalidate in-flight reads and must bump the epoch. */
+  private hadStore = false;
+
+  /**
+   * Role transitions run on their OWN chain, never on the opQueue: parked ops
+   * sit INSIDE opQueue entries waiting for the role gate, so a transition
+   * queued behind them could never run to release them (deadlock). Transitions
+   * are safe off-queue because in-flight ops on a dead transport were already
+   * rejected, parked ops only resume after the transition completes, and the
+   * worker serializes everything worker-side anyway.
+   */
+  private transitionChain: Promise<unknown> = Promise.resolve();
+
+  private chainTransition<T>(fn: () => Promise<T>): Promise<T> {
+    const result = this.transitionChain.then(fn, fn);
+    this.transitionChain = result.then(
+      () => undefined,
+      () => undefined
+    );
+    return result;
+  }
+
+  private openRoleGate(reason: string): void {
+    if (this.roleGate) return;
+    let release!: () => void;
+    let rejectFn!: (e: Error) => void;
+    const promise = new Promise<void>((res, rej) => {
+      release = res;
+      rejectFn = rej;
+    });
+    // Swallow the timeout rejection for waiters that already resolved.
+    promise.catch(() => {});
+    this.roleGate = { promise, release };
+    this.roleGateTimer = setTimeout(() => {
+      // Nothing adopted us in time: reject the parked ops (retryable) and drop
+      // the gate so later calls fail fast with 'not connected'.
+      rejectFn(new BrokerPortClosedError(`no leader adopted this tab: ${reason}`));
+      this.roleGate = null;
+      this.roleGateTimer = null;
+    }, 20_000);
+  }
+
+  private closeRoleGate(): void {
+    if (this.roleGateTimer) clearTimeout(this.roleGateTimer);
+    this.roleGateTimer = null;
+    this.roleGate?.release();
+    this.roleGate = null;
+  }
+
+  /**
+   * Become the store owner (leader). Boot and failover share this path; a
+   * failover (a previous transport existed) bumps the epoch FIRST so every
+   * in-flight chain that captured the old epoch fences itself. `resumeHeld`
+   * keeps the live worker after a broker restart and only rolls the
+   * per-leadership lock forward.
+   */
+  async adoptOwner(
+    bucketId: string,
+    opts: {
+      workerLockName: string;
+      allowMemoryFallback: boolean;
+      resumeHeld: boolean;
+    }
+  ): Promise<StorageHealth> {
+    this.roleLabel = 'leader';
+    return this.chainTransition(async () => {
+      if (
+        opts.resumeHeld &&
+        this.transport?.kind === 'worker' &&
+        this.transport.connected &&
+        this.bucketId === bucketId
+      ) {
+        await this.rawCall('relock', { workerLockName: opts.workerLockName });
+        this.closeRoleGate();
+        return this.storageHealthValue;
+      }
+      if (this.transport) await this.drainInFlight();
+      if (this.hadStore) this.storeEpoch++;
+      if (this.transport) {
+        try {
+          if (this.transport.kind === 'worker') await this.rawCall('close');
+        } catch {
+          /* ignore */
+        }
+        this.transport.close('adopting ownership', new BrokerPortClosedError('adopting ownership'));
+        this.transport = null;
+      }
+      await this.openInternal(bucketId, {
+        workerLockName: opts.workerLockName,
+        openOptions: {
+          ...PROMOTION_OPEN_OPTIONS,
+          disallowMemoryFallback: !opts.allowMemoryFallback,
+        },
+      });
+      getStats().roleChanges = (getStats().roleChanges ?? 0) + 1;
+      this.hadStore = true;
+      this.closeRoleGate();
+      return this.storageHealthValue;
+    });
+  }
+
+  /** Attach to a leader's worker through `dbPort` (follower). */
+  async adoptAttached(
+    dbPort: MessagePort,
+    snapshot: { bucketId: string; storageHealth: StorageHealth },
+    onPortDead: (reason: string) => void
+  ): Promise<void> {
+    this.roleLabel = 'follower';
+    return this.chainTransition(async () => {
+      if (this.transport) await this.drainInFlight();
+      if (this.hadStore) this.storeEpoch++;
+      this.transport?.close(
+        'adopting leader port',
+        new BrokerPortClosedError('adopting leader port')
+      );
+      this.transport = new PortSqliteTransport(dbPort, onPortDead, this.logger);
+      this.bucketId = snapshot.bucketId;
+      // The shared store exists and is seeded; mirror the owner's bookkeeping.
+      this.knownTables.clear();
+      for (const t of SYSTEM_TABLES) this.knownTables.add(t);
+      const health: StorageHealth = { ...snapshot.storageHealth, role: 'follower' };
+      this.setStorageHealth(health);
+      const stats = getStats();
+      stats.persisted = snapshot.storageHealth.status === 'persistent';
+      stats.roleChanges = (stats.roleChanges ?? 0) + 1;
+      delete stats.opfsError;
+      this.hadStore = true;
+      this.closeRoleGate();
+    });
+  }
+
+  /** Demoted while owning the store (zombie thaw, stale promotion): tear the
+   *  worker down so its OPFS handles free up, then park until re-adopted. */
+  async releaseOwnership(): Promise<void> {
+    await this.chainTransition(async () => {
+      if (this.transport) {
+        // Let ops already at the worker answer first. Without this they die
+        // with the transport — a bucket switch (moveToBucket → teardownLeader)
+        // runs while the opQueue may still have a query's fetch in flight, and
+        // that fetch's caller is not necessarily prepared to retry.
+        await this.drainInFlight();
+        try {
+          if (this.transport.kind === 'worker') {
+            await (this.transport as WorkerSqliteTransport).shutdown();
+          }
+        } catch {
+          /* worker may already be fenced/dead */
+        }
+        // Anything still pending (drain timed out) is a deliberate teardown,
+        // not a crash: fail it the way a follower's lost port does, so the
+        // error text is honest and callers can treat it as retryable.
+        this.transport.close(
+          'ownership released',
+          new BrokerPortClosedError('ownership released')
+        );
+        this.transport = null;
+      }
+      this.storeEpoch++;
+      this.openRoleGate('ownership released');
+    });
+  }
+
+  /** The leader (or its port) died. Called from the port-dead callback and the
+   *  coordinator; NOT enqueued, so in-flight ops reject immediately instead of
+   *  waiting behind whatever is stuck. */
+  onLeaderLost(reason: string): void {
+    if (this.transport?.kind === 'port') {
+      this.transport.close(reason);
+      this.transport = null;
+    }
+    this.storeEpoch++;
+    this.openRoleGate(reason);
+  }
+
+  /** Leader side: forward a follower's dbPort into the owned worker. */
+  async exposeClientPort(clientId: string, port: MessagePort): Promise<void> {
+    if (this.transport?.kind !== 'worker') {
+      throw new Error('SqliteCacheEngine: not the store owner');
+    }
+    await (this.transport as WorkerSqliteTransport).addClientPort(clientId, port);
+  }
+
+  async removeClientPort(clientId: string): Promise<void> {
+    if (this.transport?.kind !== 'worker') return;
+    await (this.transport as WorkerSqliteTransport).removeClientPort(clientId);
+  }
+
+  /** Graceful pagehide as owner: release OPFS handles NOW so the next leader
+   *  does not race the browser's worker GC. */
+  async shutdownOwnedWorker(): Promise<void> {
+    if (this.transport?.kind !== 'worker') return;
+    try {
+      await (this.transport as WorkerSqliteTransport).shutdown();
+    } catch {
+      /* ignore */
+    }
+    this.transport.close('shutdown');
+    this.transport = null;
+  }
+
+  private async ensureTable(table: string): Promise<void> {
+    if (this.knownTables.has(table)) return;
+    await this.call('run', {
+      sql: `CREATE TABLE IF NOT EXISTS "${table}" (id TEXT PRIMARY KEY, data TEXT NOT NULL)`,
+    });
+    this.knownTables.add(table);
+  }
+
+  private async execRows(sql: string, bind: unknown[]): Promise<Row[]> {
+    const { rows } = await this.call<{ rows: { data: string }[] }>('exec', { sql, bind });
+    const s = getStats();
+    const t0 = performance.now();
+    const out = (rows ?? []).map((r) => {
+      s.bytesParsed += r.data.length;
+      return reviveRow(r.data);
+    });
+    s.parseMs += performance.now() - t0;
+    s.rowsParsed += out.length;
+    return out;
+  }
+
+  // ---- reads ---------------------------------------------------------------
+
+  async select(plan: QueryPlan, params: Record<string, unknown> = {}): Promise<Row[]> {
+    if (this.workerSelect) {
+      // ONE round-trip: the worker executes the whole plan (table creation,
+      // base select, relation tree, JSON parse) and returns structured-clone
+      // rows. The legacy path below pays a postMessage hop per table/relation
+      // level plus main-thread parsing — the dominant first-load cost.
+      //
+      // Normalize before postMessage: class-instance VALUES (RecordId & co) →
+      // their `stableKey` string. structuredClone strips a class instance to a
+      // bare plain object — and surrealdb's RecordId keeps its data behind
+      // getters (no own properties), so it clones to `{}`: the worker would
+      // filter on garbage. Applies to params AND to values baked inside the
+      // plan (where nodes, relation sub-wheres, window ids).
+      // Param KEYS must pass through untouched — `comparisonSql` resolves
+      // `paramRef` via hasOwnProperty, and a dropped key silently falls back to
+      // the baked literal (the crossed-results class fixed in aa4af79b).
+      const normPlan = normalizePlanForClone(plan);
+      const normParams: Record<string, unknown> = {};
+      for (const [k, v] of Object.entries(params)) normParams[k] = toCloneSafe(v);
+      try {
+        const res = await this.call<{ rows: Row[]; relationFetches?: number }>('select', {
+          plan: normPlan,
+          params: normParams,
+        });
+        getStats().relationFetches += res.relationFetches ?? 0;
+        return res.rows ?? [];
+      } catch (err) {
+        // Stale worker script without the 'select' op: fall back for good.
+        if (err instanceof Error && err.message.includes('unknown message')) {
+          this.workerSelect = false;
+          this.logger.warn(
+            { err, Category: 'sp00ky-client::SqliteCacheEngine::select' },
+            'Worker lacks select op (stale script?) — falling back to multi-hop select'
+          );
+        } else {
+          throw err;
+        }
+      }
+    }
+    return this.selectLegacy(plan, params);
+  }
+
+  /** Pre-worker-select path: one worker round-trip per table/relation level,
+   *  rows parsed on the main thread. Kept as the `workerSelect:false` escape
+   *  hatch and the stale-worker fallback. */
+  private async selectLegacy(
+    plan: QueryPlan,
+    params: Record<string, unknown> = {}
+  ): Promise<Row[]> {
+    // Window materialization: base rows are exactly `plan.ids`, ordered.
+    if (plan.ids) {
+      const result = await this.selectByIds(plan.table, plan.ids, {
+        select: plan.select,
+        orderBy: plan.orderBy,
+      });
+      await resolveRelations(result, plan.relations, this);
+      return result;
+    }
+    await this.ensureTable(plan.table);
+    const bind: unknown[] = [];
+    // Projection binds land in the SELECT clause, ahead of any WHERE binds.
+    const proj = plan.select ? projectedDataSql(plan.select, bind) : 'data';
+    let sql = `SELECT ${proj} FROM "${plan.table}"`;
+    if (plan.where && plan.where.length > 0) {
+      sql += ` WHERE ${renderWhereSql(plan.where, bind, params)}`;
+    }
+    if (plan.orderBy && plan.orderBy.length > 0) sql += renderOrderSql(plan.orderBy);
+    // Deterministic fallback, in parity with `sqlite-select.ts`: without it an
+    // unordered query renders in insertion order here and in membership order
+    // after the server answers, which reshuffles the list on screen.
+    else sql += ` ORDER BY id`;
+    if (plan.limit !== undefined) sql += ` LIMIT ${Number(plan.limit)}`;
+    if (plan.offset !== undefined) sql += ` OFFSET ${Number(plan.offset)}`;
+    const rows = await this.execRows(sql, bind);
+    await resolveRelations(rows, plan.relations, this);
+    return rows;
+  }
+
+  async fetchRelation(req: RelationFetch): Promise<Row[]> {
+    getStats().relationFetches++;
+    await this.ensureTable(req.table);
+    const keys = req.keys.map(stableKey);
+    const placeholders = keys.map(() => '?').join(', ');
+    const bind: unknown[] = [...keys];
+    const lhs = req.matchField === 'id' ? 'id' : `json_extract(data, '$.${req.matchField}')`;
+    let sql = `SELECT data FROM "${req.table}" WHERE ${lhs} IN (${placeholders})`;
+    if (req.where && req.where.length > 0) {
+      sql += ` AND ${renderWhereSql(req.where, bind, {})}`;
+    }
+    if (req.orderBy && req.orderBy.length > 0) sql += renderOrderSql(req.orderBy);
+    const rows = await this.execRows(sql, bind);
+    return req.select ? rows.map((r) => project(r, req.select!)) : rows;
+  }
+
+  async selectByIds(
+    table: string,
+    ids: Id[],
+    opts?: { select?: string[]; orderBy?: OrderBy }
+  ): Promise<Row[]> {
+    if (ids.length === 0) return [];
+    await this.ensureTable(table);
+    const keys = ids.map(stableKey);
+    const placeholders = keys.map(() => '?').join(', ');
+    // Projection binds land in the SELECT clause, so they go first. Must stay
+    // byte-identical to the worker path in sqlite-select.ts.
+    const bind: unknown[] = [];
+    const dataCol = opts?.select ? projectedDataSql(opts.select, bind) : 'data';
+    bind.push(...keys);
+    let sql = `SELECT ${dataCol} FROM "${table}" WHERE id IN (${placeholders})`;
+    if (opts?.orderBy && opts.orderBy.length > 0) sql += renderOrderSql(opts.orderBy);
+    let rows = await this.execRows(sql, bind);
+    if (!opts?.orderBy || opts.orderBy.length === 0) {
+      const pos = new Map(keys.map((k, i) => [k, i]));
+      rows = rows.sort((a, b) => (pos.get(stableKey(a.id)) ?? 0) - (pos.get(stableKey(b.id)) ?? 0));
+    }
+    return rows;
+  }
+
+  async getById(table: string, id: Id): Promise<Row | null> {
+    await this.ensureTable(table);
+    const rows = await this.execRows(`SELECT data FROM "${table}" WHERE id = ?`, [stableKey(id)]);
+    return rows[0] ?? null;
+  }
+
+  /**
+   * `(id, _00_rv)` of every row per table, straight off the worker: no body
+   * parse, one round trip per table. A table that does not exist yet reads as
+   * empty rather than failing the whole scan.
+   */
+  async scanVersions(tables: string[]): Promise<Record<string, [string, number][]>> {
+    const out: Record<string, [string, number][]> = {};
+    for (const table of tables) {
+      try {
+        const { rows } = await this.call<{ rows: { id: string; rv: unknown }[] }>('exec', {
+          sql: `SELECT id, json_extract(data, '$._00_rv') AS rv FROM "${table.replace(/"/g, '""')}"`,
+        });
+        out[table] = (rows ?? []).map((r) => [r.id, Number(r.rv) || 0]);
+      } catch {
+        out[table] = [];
+      }
+    }
+    return out;
+  }
+
+  async getSnapshot(key: string): Promise<StoredSnapshot | null> {
+    const { rows } = await this.call<{ rows: { id: string; data: unknown }[] }>('exec', {
+      sql: `SELECT id, data FROM "${SNAPSHOT_TABLE}" WHERE id IN (?, ?)`,
+      bind: [key, `${key}:meta`],
+    });
+    let bytes: Uint8Array | null = null;
+    let meta: SnapshotMeta | null = null;
+    for (const r of rows ?? []) {
+      if (r.id === key) {
+        if (r.data instanceof Uint8Array) bytes = r.data;
+        else if (r.data instanceof ArrayBuffer) bytes = new Uint8Array(r.data);
+        else if (typeof r.data === 'string') bytes = new TextEncoder().encode(r.data);
+      } else if (typeof r.data === 'string') {
+        try {
+          meta = JSON.parse(r.data) as SnapshotMeta;
+        } catch {
+          meta = null;
+        }
+      }
+    }
+    return bytes && meta ? { bytes, meta } : null;
+  }
+
+  async putSnapshot(key: string, bytes: Uint8Array, meta: SnapshotMeta): Promise<void> {
+    // One transaction, so a reader never sees new bytes with old meta.
+    await this.call('batch', [
+      {
+        sql: `INSERT OR REPLACE INTO "${SNAPSHOT_TABLE}" (id, data) VALUES (?, ?)`,
+        bind: [key, bytes],
+      },
+      {
+        sql: `INSERT OR REPLACE INTO "${SNAPSHOT_TABLE}" (id, data) VALUES (?, ?)`,
+        bind: [`${key}:meta`, JSON.stringify(meta)],
+      },
+    ]);
+  }
+
+  async deleteSnapshot(key: string): Promise<void> {
+    await this.call('run', {
+      sql: `DELETE FROM "${SNAPSHOT_TABLE}" WHERE id IN (?, ?)`,
+      bind: [key, `${key}:meta`],
+    });
+  }
+
+  // ---- writes --------------------------------------------------------------
+
+  async upsert(table: string, id: Id, data: Row, mode: 'replace' | 'merge'): Promise<void> {
+    await this.ensureTable(table);
+    const key = stableKey(id);
+    if (mode === 'merge') {
+      // Merge in-SQL via json_patch (RFC7396 = MERGE semantics): on insert store
+      // the row, on conflict shallow-merge. Serialize once, reuse for VALUES and
+      // the patch. No read-modify-write round-trip. (RFC7396: null deletes key,
+      // which is how a key holding `undefined` clears the stored field.)
+      const { insert, patch } = serializeMergeRow({ ...data, id: key });
+      await this.call('run', {
+        sql: `INSERT INTO "${table}"(id, data) VALUES(?, ?) ON CONFLICT(id) DO UPDATE SET data = json_patch(data, ?)`,
+        bind: [key, insert, patch],
+      });
+      return;
+    }
+    await this.call('run', {
+      sql: `INSERT INTO "${table}"(id, data) VALUES(?, ?) ON CONFLICT(id) DO UPDATE SET data = excluded.data`,
+      bind: [key, serializeRow({ ...data, id: key })],
+    });
+  }
+
+  async patch(table: string, id: Id, patches: unknown[]): Promise<void> {
+    // RFC6902 (fast-json-patch) applied read-modify-write — SQLite's json_patch
+    // is RFC7396 merge-patch and would misinterpret op arrays.
+    const existing = (await this.getById(table, id)) ?? { id: stableKey(id) };
+    const next = applyPatch(existing, patches as Operation[]).newDocument as Row;
+    await this.upsert(table, id, next, 'replace');
+  }
+
+  async delete(table: string, id: Id): Promise<void> {
+    await this.ensureTable(table);
+    await this.call('run', { sql: `DELETE FROM "${table}" WHERE id = ?`, bind: [stableKey(id)] });
+  }
+
+  // ---- SurrealQL-vocabulary shim (LocalStore compatibility) ---------------
+
+  /**
+   * Execute a raw SurrealQL statement by translating the client's bounded
+   * vocabulary to verbs (see `surql-translate.ts`). Returns results shaped like
+   * SurrealDB's `.query()` (one element per statement; a tx prepends a `null`
+   * begin-result so `surql.seal` extraction lines up). Epoch-fences writes.
+   */
+  async query<T extends unknown[]>(
+    sql: string,
+    vars: Record<string, unknown> = {},
+    opts?: { epoch?: number }
+  ): Promise<T> {
+    if (opts?.epoch !== undefined && opts.epoch !== this.storeEpoch) throw new StaleEpochError();
+    const start = performance.now();
+    try {
+      const { transaction, ops } = translateSurql(sql, vars);
+      let shaped: unknown[];
+      // FAST PATH: a pure-write transaction (bulk sync-down is one
+      // `tx([upsertMerge…])`) compiles to a SINGLE worker `batch` message run in
+      // one SQLite transaction — instead of 1-2 worker round-trips PER row. This
+      // is the dominant sync-down cost; per-op execution here caused the churn
+      // OOM. Mixed txs (LET/RETURN single-record mutations) keep the per-op path.
+      if (
+        transaction &&
+        ops.every(
+          (o) =>
+            o.kind === 'upsert' ||
+            o.kind === 'delete' ||
+            o.kind === 'deleteAll' ||
+            o.kind === 'noop'
+        )
+      ) {
+        await this.runWriteBatch(ops);
+        // Shape each statement's result the SAME as the per-op path (`execOp`),
+        // so a caller reading a statement's output still works after taking the
+        // batch fast path. A single `create()` compiles to an all-upsert tx
+        // (createSet + createMutation) and reads `resultIndex:0` for the new
+        // row + its id; returning empty arrays here dropped that row (id became
+        // undefined → the reconcile crashed in `encodeRecordId`). The
+        // single-batch write is kept — this only rebuilds the return value.
+        shaped = [null, ...ops.map(pureWriteOpResult)];
+      } else {
+        const results: unknown[] = [];
+        // Per-query scope holds `LET $var = (...)` bindings for later statements
+        // (e.g. `RETURN { target: $updated }`).
+        const scope: Record<string, unknown> = {};
+        for (const op of ops) results.push(await this.execOp(op, scope, vars));
+        shaped = transaction ? [null, ...results] : results;
+      }
+      this.events.emit(DatabaseEventTypes.LocalQuery, {
+        query: sql,
+        vars,
+        duration: performance.now() - start,
+        success: true,
+        timestamp: Date.now(),
+      });
+      return shaped as unknown as T;
+    } catch (err) {
+      this.events.emit(DatabaseEventTypes.LocalQuery, {
+        query: sql,
+        vars,
+        duration: performance.now() - start,
+        success: false,
+        error: err instanceof Error ? err.message : String(err),
+        timestamp: Date.now(),
+      });
+      throw err;
+    }
+  }
+
+  async execute<R>(
+    query: SealedQuery<R>,
+    vars?: Record<string, unknown>,
+    opts?: { epoch?: number }
+  ): Promise<R> {
+    const raw = await this.query<unknown[]>(query.sql, vars, opts);
+    return query.extract(raw);
+  }
+
+  queryUngated<T extends unknown[]>(sql: string, vars?: Record<string, unknown>): Promise<T> {
+    return this.query<T>(sql, vars);
+  }
+
+  private async execOp(
+    op: SqlOp,
+    scope: Record<string, unknown>,
+    vars: Record<string, unknown>
+  ): Promise<unknown> {
+    switch (op.kind) {
+      case 'getById': {
+        const row = await this.getById(tableOf(op.id), op.id);
+        if (op.value) return row ? (row[op.value] ?? null) : null;
+        return row ? (op.select ? project(row, op.select) : row) : null;
+      }
+      case 'selectByIds': {
+        if (op.ids.length === 0) return [];
+        let rows = await this.selectByIds(tableOf(op.ids[0]), op.ids, {
+          select: op.select,
+          orderBy: op.orderBy,
+        });
+        // Windowed in JS, not SQL: the id list is already the whole result set,
+        // and its order is restored after the fetch (see selectByIds).
+        if (op.start !== undefined || op.limit !== undefined) {
+          const from = op.start ?? 0;
+          rows = rows.slice(from, op.limit === undefined ? undefined : from + op.limit);
+        }
+        if (op.value) return rows.map((r) => r[op.value!]);
+        return rows;
+      }
+      case 'selectTable': {
+        const rows = await this.rawSelectTable(op.table, op.where, op.orderBy, {
+          limit: op.limit,
+          start: op.start,
+        });
+        if (op.value) return rows.map((r) => r[op.value!]);
+        return op.select ? rows.map((r) => project(r, op.select!)) : rows;
+      }
+      case 'count': {
+        await this.ensureTable(op.table);
+        const bind: unknown[] = [];
+        let sql = `SELECT COUNT(*) AS n FROM "${op.table}"`;
+        if (op.where && op.where.length > 0) sql += ` WHERE ${renderWhereSql(op.where, bind, {})}`;
+        const { rows } = await this.call<{ rows: { n: number }[] }>('exec', { sql, bind });
+        // `GROUP ALL` collapses to a single `{ count }` row on SurrealDB; match
+        // it exactly so callers can read `rows[0].count` on either engine.
+        return [{ count: rows?.[0]?.n ?? 0 }];
+      }
+      case 'infoForDb': {
+        const { rows } = await this.call<{ rows: { name: string }[] }>('exec', {
+          sql: "SELECT name FROM sqlite_master WHERE type = 'table' AND name NOT LIKE 'sqlite_%' ORDER BY name",
+        });
+        // SurrealDB answers with the DEFINE statement per table. Nothing here
+        // parses it (the DevTools explorer only reads the keys), so a truthful
+        // schemaless stand-in beats inventing field definitions.
+        const tables: Record<string, string> = {};
+        for (const r of rows ?? []) tables[r.name] = `DEFINE TABLE ${r.name} SCHEMALESS`;
+        return { tables, analyzers: {}, functions: {}, params: {}, users: {} };
+      }
+      case 'upsert':
+        await this.upsert(tableOf(op.id), op.id, op.data, op.mode);
+        // Cheap return — no read-back. The full merged row is only needed by a
+        // LET-wrapped upsert, which reads it back in the 'let' case below. This
+        // avoids an extra worker round-trip + full-row parse on EVERY sync-down
+        // write (the hot path under rapid churn). Shared with the batch fast
+        // path so the two never drift.
+        return pureWriteOpResult(op);
+      case 'updateSet': {
+        const existing = (await this.getById(tableOf(op.id), op.id)) ?? { id: stableKey(op.id) };
+        for (const { path, op: setOp, value } of op.sets) {
+          if (setOp === '+=' || setOp === '-=') {
+            const cur = Number(getPath(existing, path) ?? 0);
+            const delta = Number(value ?? 0);
+            setPath(existing, path, setOp === '+=' ? cur + delta : cur - delta);
+          } else {
+            setPath(existing, path, value);
+          }
+        }
+        await this.upsert(tableOf(op.id), op.id, existing, 'replace');
+        return op.returnNone ? null : existing;
+      }
+      case 'delete':
+        await this.delete(tableOf(op.id), op.id);
+        return pureWriteOpResult(op);
+      case 'deleteAll':
+        await this.ensureTable(op.table);
+        await this.call('run', { sql: `DELETE FROM "${op.table}"` });
+        return pureWriteOpResult(op);
+      case 'let': {
+        let result = await this.execOp(op.inner, scope, vars);
+        // A LET-bound UPSERT must expose the FULL merged row (e.g.
+        // `RETURN { target: $updated }`), so read it back here — only here,
+        // not on every upsert.
+        if (op.inner.kind === 'upsert') {
+          result = (await this.getById(tableOf(op.inner.id), op.inner.id)) ?? result;
+        }
+        scope[op.var] = result;
+        return result;
+      }
+      case 'return': {
+        const obj: Row = {};
+        for (const { key, var: v } of op.entries) {
+          obj[key] = v in scope ? scope[v] : vars[v];
+        }
+        return obj;
+      }
+      case 'noop':
+        return pureWriteOpResult(op);
+    }
+  }
+
+  /**
+   * Compile a pure-write op list to ONE worker `batch` (single SQLite
+   * transaction). Ensures each touched table exists, then one statement per op.
+   * Merges happen in-SQL via json_patch — no read-back round-trips.
+   */
+  private async runWriteBatch(ops: SqlOp[]): Promise<void> {
+    const stmts: { sql: string; bind?: unknown[] }[] = [];
+    const tables = new Set<string>();
+    const ensure = (t: string) => {
+      if (!tables.has(t)) {
+        tables.add(t);
+        this.knownTables.add(t);
+        stmts.push({
+          sql: `CREATE TABLE IF NOT EXISTS "${t}" (id TEXT PRIMARY KEY, data TEXT NOT NULL)`,
+        });
+      }
+    };
+    for (const op of ops) {
+      if (op.kind === 'upsert') {
+        const t = tableOf(op.id);
+        const key = stableKey(op.id);
+        ensure(t);
+        if (op.mode === 'merge') {
+          // Serialize ONCE and reuse for both VALUES (fresh insert) and the
+          // json_patch (merge). Patching with `id` is a harmless no-op set, so
+          // the full row doubles as the delta — halves per-row stringify cost.
+          // A key holding `undefined` (a field cleared upstream) rides in the
+          // delta as a `null` member, which json_patch removes.
+          const { insert, patch } = serializeMergeRow({ ...op.data, id: key });
+          stmts.push({
+            sql: `INSERT INTO "${t}"(id, data) VALUES(?, ?) ON CONFLICT(id) DO UPDATE SET data = json_patch(data, ?)`,
+            bind: [key, insert, patch],
+          });
+        } else {
+          stmts.push({
+            sql: `INSERT INTO "${t}"(id, data) VALUES(?, ?) ON CONFLICT(id) DO UPDATE SET data = excluded.data`,
+            bind: [key, serializeRow({ ...op.data, id: key })],
+          });
+        }
+      } else if (op.kind === 'delete') {
+        const t = tableOf(op.id);
+        ensure(t);
+        stmts.push({ sql: `DELETE FROM "${t}" WHERE id = ?`, bind: [stableKey(op.id)] });
+      } else if (op.kind === 'deleteAll') {
+        ensure(op.table);
+        stmts.push({ sql: `DELETE FROM "${op.table}"` });
+      }
+    }
+    if (stmts.length > 0) await this.call('batch', stmts);
+  }
+
+  private async rawSelectTable(
+    table: string,
+    where?: WhereNode[],
+    orderBy?: OrderBy,
+    window?: { limit?: number; start?: number }
+  ): Promise<Row[]> {
+    await this.ensureTable(table);
+    const bind: unknown[] = [];
+    let sql = `SELECT data FROM "${table}"`;
+    if (where && where.length > 0) sql += ` WHERE ${renderWhereSql(where, bind, {})}`;
+    if (orderBy && orderBy.length > 0) sql += renderOrderSql(orderBy);
+    // SQLite has no bare OFFSET, so a START without a LIMIT needs `LIMIT -1`
+    // (its documented "no limit" sentinel) to stay valid SQL.
+    if (window?.limit !== undefined) sql += ` LIMIT ${Number(window.limit)}`;
+    else if (window?.start !== undefined) sql += ' LIMIT -1';
+    if (window?.start !== undefined) sql += ` OFFSET ${Number(window.start)}`;
+    return this.execRows(sql, bind);
+  }
+
+  async transaction<T>(fn: (tx: EngineTx) => Promise<T>): Promise<T> {
+    // Verbs run in order on the single worker message channel. `patch` needs a
+    // read-modify-write round-trip, so a single BEGIN/COMMIT batch cannot wrap
+    // the whole closure; sequential execution is sufficient for the current
+    // single-record write sites.
+    const tx: EngineTx = {
+      upsert: (t, id, data, mode) => this.upsert(t, id, data, mode),
+      patch: (t, id, p) => this.patch(t, id, p),
+      delete: (t, id) => this.delete(t, id),
+    };
+    return fn(tx);
+  }
+}
+
+// ==================== instrumentation ====================
+
+interface SqliteStats {
+  roundTrips: number;
+  batchStatements: number;
+  maxBatch: number;
+  inFlight: number;
+  maxInFlight: number;
+  byType: Record<string, number>;
+  /** Worker round trips that hit `localOpTimeoutMs` (see rawCall). */
+  timeouts?: number;
+  /** Time ops spent waiting behind the opQueue before dispatch. */
+  queueWaitMs: number;
+  /** Time inside the worker's message handler (actual SQLite work). */
+  workerMs: number;
+  /** Round-trip time minus workerMs: postMessage + scheduling overhead. */
+  rpcOverheadMs: number;
+  /** Main-thread JSON parse/revive of returned rows. */
+  parseMs: number;
+  rowsParsed: number;
+  bytesParsed: number;
+  /** Relation-resolver fan-out fetches (one worker round-trip each). */
+  relationFetches: number;
+  /** Whether the open store is OPFS-backed. `false` here with an `opfsError`
+   *  means the whole dataset is sitting in RAM. Optional so it stays absent
+   *  until the first open (and is skipped by the backfill loop below). */
+  persisted?: boolean;
+  /** Why OPFS persistence failed, when it did. */
+  opfsError?: string;
+  /** Shared-tabs follower: ops that crossed the leader's MessagePort. */
+  proxiedOps?: number;
+  /** Times this tab's engine changed hands (promotions + attachments). */
+  roleChanges?: number;
+}
+
+const EMPTY_STATS: SqliteStats = {
+  roundTrips: 0,
+  batchStatements: 0,
+  maxBatch: 0,
+  inFlight: 0,
+  maxInFlight: 0,
+  byType: {},
+  queueWaitMs: 0,
+  workerMs: 0,
+  rpcOverheadMs: 0,
+  parseMs: 0,
+  rowsParsed: 0,
+  bytesParsed: 0,
+  relationFetches: 0,
+};
+
+/** Live stats, inspectable in the browser console via `__sqliteStats`. Counts
+ *  worker round-trips (the sync-down cost driver), batch sizes, queue depth
+ *  (`maxInFlight`), and the latency split of each round-trip (queue wait vs
+ *  worker time vs RPC overhead vs main-thread row parsing) so first-load cost
+ *  can be measured rather than guessed. */
+function getStats(): SqliteStats {
+  const g = globalThis as unknown as { __sqliteStats?: SqliteStats };
+  if (!g.__sqliteStats) {
+    g.__sqliteStats = { ...EMPTY_STATS, byType: {} };
+  } else {
+    // Backfill fields added since the object was created (HMR / older bundle).
+    for (const [k, v] of Object.entries(EMPTY_STATS)) {
+      if ((g.__sqliteStats as any)[k] === undefined) (g.__sqliteStats as any)[k] = v;
+    }
+  }
+  return g.__sqliteStats;
+}
+
+// SQL rendering + row (de)serialization live in `sqlite-plan-sql.ts`, shared
+// with the worker (worker-side plan execution renders the same SQL).
+
+// ==================== structured-clone normalization ====================
+
+/**
+ * Make a bind/param value safe to cross the worker boundary. structuredClone
+ * keeps plain data intact but strips a CLASS instance to a bare object — and
+ * surrealdb's `RecordId` stores its fields behind getters (zero own
+ * properties), so it clones to `{}`. Convert such instances to their
+ * `stableKey` string (the exact value `scalar()` would bind on the main
+ * thread), leave everything clone-representable untouched.
+ */
+function toCloneSafe(v: unknown): unknown {
+  if (v === null || typeof v !== 'object') return v;
+  if (Array.isArray(v) || v instanceof Uint8Array || v instanceof Date) return v;
+  const proto = Object.getPrototypeOf(v);
+  if (proto === Object.prototype || proto === null) return v;
+  return stableKey(v);
+}
+
+function normalizeWhereForClone(nodes: WhereNode[] | undefined): WhereNode[] | undefined {
+  if (!nodes) return nodes;
+  return nodes.map((n) =>
+    'or' in n
+      ? { or: n.or.map((c) => ({ ...c, value: toCloneSafe(c.value) })) }
+      : { ...n, value: toCloneSafe(n.value) }
+  );
+}
+
+function normalizeRelationForClone(r: RelationPlan): RelationPlan {
+  return {
+    ...r,
+    where: normalizeWhereForClone(r.where),
+    relations: r.relations?.map(normalizeRelationForClone),
+  };
+}
+
+/** Normalize every baked value a plan carries (where trees, window ids) for
+ *  the postMessage to the worker's `select` op. */
+function normalizePlanForClone(plan: QueryPlan): QueryPlan {
+  return {
+    ...plan,
+    ids: plan.ids?.map(stableKey),
+    where: normalizeWhereForClone(plan.where),
+    relations: plan.relations?.map(normalizeRelationForClone),
+  };
+}