diff --git a/app/(main)/[locale]/page.tsx b/app/(main)/[locale]/page.tsx
index 3d8250b9..74843917 100644
--- a/app/(main)/[locale]/page.tsx
+++ b/app/(main)/[locale]/page.tsx
@@ -1080,6 +1080,21 @@ export default function Home() {
         } catch {
           /* the index is optional */
         }
+        // The offline REPLICA's launch catch-up. Same reasoning as the index's,
+        // plus one of its own: a `/changes` cursor cannot tell us about anything
+        // that happened while the process was dead, so a cycle at launch is what
+        // drains the backlog. One cycle is bounded, so a first sync of a large
+        // mailbox needs several - `chainSync` runs them with a hard cap.
+        //
+        // Sequenced AFTER the index rather than in parallel: both write the same
+        // SQLite file, and although `busy_timeout` makes concurrent writers safe,
+        // there is no reason to spend the contention during first paint.
+        try {
+          const { chainSync } = await import('@/lib/offline-replica-client');
+          await chainSync({ slot: useAccountStore.getState().getActiveAccount()?.cookieSlot });
+        } catch {
+          /* the replica is optional */
+        }
       })();
       // Deliberately after the initial mailbox fetch settles: the catch-up is a
       // background nicety and must not compete with first paint.
diff --git a/app/api/offline/mail/route.ts b/app/api/offline/mail/route.ts
new file mode 100644
index 00000000..b5bc1124
--- /dev/null
+++ b/app/api/offline/mail/route.ts
@@ -0,0 +1,87 @@
+// GET /api/offline/mail?kind=mailboxes|list|message - the OFFLINE READ SURFACE.
+//
+// THIS ROUTE MUST NEVER MAKE A NETWORK CALL. That is the whole feature: it is
+// consulted precisely when the backend is unreachable, so a JMAP session fetch to
+// learn the account id would fail for the exact reason the route was called. The
+// account is resolved from the request's own encrypted `jmap_stalwart_ctx` cookie
+// (a local decrypt) and from the account ids the store already holds rows for.
+//
+// It is a FALLBACK, not a cache in front of the server - see `read.ts`'s header
+// for the coherence rules that depend on that, and `lib/offline-fallback-client.ts`
+// for the one place that decides a read has genuinely failed.
+import { NextRequest, NextResponse } from 'next/server';
+import {
+  readEnvelopePage, readMailboxes, readMessage,
+} from '@/lib/offline-replica/read';
+import {
+  resolveIndexSession, resolveReadAccountId, withReplica,
+} from '@/lib/offline-replica/engine';
+import { gateReplicaRoute, NO_STORE, replicaErrorResponse } from '@/lib/offline-replica/route-gate';
+
+export const runtime = 'nodejs';
+export const dynamic = 'force-dynamic';
+
+const MAX_LIMIT = 200;
+
+export async function GET(request: NextRequest) {
+  const gated = gateReplicaRoute();
+  if (gated) return gated;
+
+  const params = request.nextUrl.searchParams;
+  const kind = params.get('kind') ?? 'mailboxes';
+  if (kind !== 'mailboxes' && kind !== 'list' && kind !== 'message') {
+    return NextResponse.json({ error: 'kind must be mailboxes, list or message' }, { status: 400 });
+  }
+
+  try {
+    const session = await resolveIndexSession(request);
+    const payload = await withReplica(session.accountId, (store) => {
+      const jmapAccountId = resolveReadAccountId(store, params.get('jmapAccountId'));
+      if (!jmapAccountId) {
+        // Nothing synced yet for this account. Not an error - the caller falls
+        // back to whatever it would have shown without a replica.
+        return { empty: true as const };
+      }
+
+      if (kind === 'mailboxes') {
+        return { empty: false as const, jmapAccountId, mailboxes: readMailboxes(store, jmapAccountId) };
+      }
+
+      if (kind === 'list') {
+        const rawLimit = Number(params.get('limit') ?? '50');
+        const limit = Number.isFinite(rawLimit)
+          ? Math.min(Math.max(Math.trunc(rawLimit), 1), MAX_LIMIT)
+          : 50;
+        const rawOffset = Number(params.get('offset') ?? '0');
+        const offset = Number.isFinite(rawOffset) ? Math.max(Math.trunc(rawOffset), 0) : 0;
+        // An absent mailboxId means "everything", which is what the unified views
+        // ask for; an empty string is a caller bug and must not silently widen.
+        const mailboxParam = params.get('mailboxId');
+        const mailboxId = mailboxParam === null ? null : mailboxParam;
+        if (mailboxId === '') {
+          return { empty: true as const };
+        }
+        const page = readEnvelopePage(store, jmapAccountId, mailboxId, limit, offset);
+        return { empty: false as const, jmapAccountId, ...page };
+      }
+
+      const id = params.get('id');
+      if (!id || id.length > 256) return { empty: true as const };
+      const message = readMessage(store, jmapAccountId, id);
+      if (!message) return { empty: false as const, jmapAccountId, email: null, hasBody: false };
+      return {
+        empty: false as const,
+        jmapAccountId,
+        email: message.email,
+        hasBody: message.hasBody,
+      };
+    });
+
+    if (payload.empty) {
+      return NextResponse.json({ ok: true, available: false }, { headers: NO_STORE });
+    }
+    return NextResponse.json({ ok: true, available: true, ...payload }, { headers: NO_STORE });
+  } catch (error) {
+    return replicaErrorResponse(error, 'offline read');
+  }
+}
diff --git a/app/api/offline/status/route.ts b/app/api/offline/status/route.ts
new file mode 100644
index 00000000..7c493e43
--- /dev/null
+++ b/app/api/offline/status/route.ts
@@ -0,0 +1,104 @@
+// GET    /api/offline/status - size, freshness and retention policy, for Settings.
+// PUT    /api/offline/status - update the retention policy.
+// DELETE /api/offline/status - purge the replica.
+//
+// The POLICY LIVES IN THE ENCRYPTED STORE, not in renderer localStorage. The
+// design review's H1 was that a server-side engine cannot read a renderer-only
+// setting; keeping the policy server-side means the retention pass always has the
+// value it needs, while the DECISION TO SYNC AT ALL stays with the renderer, so
+// nothing is ever materialised for an account that never opted in.
+//
+// Like every read here, GET makes no network call: an offline user must still be
+// able to see what they have and free the space.
+import { NextRequest, NextResponse } from 'next/server';
+import { clampPolicy, POLICY_LIMITS, type RetentionPolicy } from '@/lib/offline-replica/store';
+import { resolveIndexSession, resolveReadAccountId, withReplica } from '@/lib/offline-replica/engine';
+import { gateReplicaRoute, NO_STORE, replicaErrorResponse } from '@/lib/offline-replica/route-gate';
+
+export const runtime = 'nodejs';
+export const dynamic = 'force-dynamic';
+
+export async function GET(request: NextRequest) {
+  const gated = gateReplicaRoute();
+  if (gated) return gated;
+  try {
+    const session = await resolveIndexSession(request);
+    const payload = await withReplica(session.accountId, (store) => {
+      const jmapAccountId = resolveReadAccountId(store, null);
+      const policy = store.getPolicy();
+      const flags = store.getFlags(Date.now());
+      if (!jmapAccountId) {
+        return {
+          policy,
+          limits: POLICY_LIMITS,
+          synced: false,
+          stats: null,
+          coveragePhase: 'never-run',
+          resyncRequired: flags.resyncRequired,
+          lastCycleAt: flags.lastCycleAt ?? null,
+          lastCycleOk: flags.lastCycleOk ?? null,
+        };
+      }
+      return {
+        policy,
+        limits: POLICY_LIMITS,
+        synced: true,
+        stats: store.stats(jmapAccountId),
+        coveragePhase: store.getCoverage(jmapAccountId)?.phase ?? 'never-run',
+        coveredFrom: store.getCoverage(jmapAccountId)?.coveredFrom ?? null,
+        resyncRequired: flags.resyncRequired,
+        lastCycleAt: flags.lastCycleAt ?? null,
+        lastCycleOk: flags.lastCycleOk ?? null,
+        lastCycleError: flags.lastCycleError ?? null,
+      };
+    });
+    return NextResponse.json({ ok: true, ...payload }, { headers: NO_STORE });
+  } catch (error) {
+    return replicaErrorResponse(error, 'offline status');
+  }
+}
+
+export async function PUT(request: NextRequest) {
+  const gated = gateReplicaRoute();
+  if (gated) return gated;
+
+  let body: Record<string, unknown> = {};
+  try {
+    const text = await request.text();
+    if (text.trim()) body = JSON.parse(text) as Record<string, unknown>;
+  } catch {
+    return NextResponse.json({ error: 'Malformed JSON body' }, { status: 400 });
+  }
+
+  const policy = clampPolicy(body as Partial<RetentionPolicy>);
+  try {
+    const session = await resolveIndexSession(request);
+    await withReplica(session.accountId, (store) => {
+      store.transaction(() => { store.setPolicy(policy); });
+    });
+    // The cycle applies it: a widen re-enters coverage scanning, a narrow evicts,
+    // and the clock guard is told this was INTENT rather than a glitch by the
+    // `lastEnvelopeDays` it compares against.
+    return NextResponse.json({ ok: true, policy }, { headers: NO_STORE });
+  } catch (error) {
+    return replicaErrorResponse(error, 'offline policy update');
+  }
+}
+
+export async function DELETE(request: NextRequest) {
+  const gated = gateReplicaRoute();
+  if (gated) return gated;
+  try {
+    const session = await resolveIndexSession(request);
+    await withReplica(session.accountId, (store) => {
+      // ALL OF IT, cursors included. A record wipe that leaves cursors behind is
+      // the one state no amount of syncing repairs: `/changes` structurally cannot
+      // re-deliver mail that already existed when the cursor was captured, so the
+      // next cycle would advance a live cursor over an empty store forever.
+      store.transaction(() => { store.purgeAll(); });
+    });
+    return NextResponse.json({ ok: true, purged: true }, { headers: NO_STORE });
+  } catch (error) {
+    return replicaErrorResponse(error, 'offline purge');
+  }
+}
diff --git a/app/api/offline/sync/route.ts b/app/api/offline/sync/route.ts
new file mode 100644
index 00000000..de895463
--- /dev/null
+++ b/app/api/offline/sync/route.ts
@@ -0,0 +1,49 @@
+// POST /api/offline/sync - run ONE bounded delta-sync cycle for the calling
+// session's account.
+//
+// The renderer drives this: once at launch (catch-up for whatever changed while
+// the app was closed, for which no push event was ever delivered) and on each
+// JMAP `StateChange` from the live push connection. There is no background worker
+// and no resident credential - see `lib/offline-replica/sync.ts`'s header for why
+// that architecture choice keeps most of the original design review's critical
+// findings out of scope entirely.
+//
+// A cycle is BOUNDED. `unfinishedWork: true` means "call again", and the renderer
+// chains with a cap; it never means an error.
+import { NextRequest, NextResponse } from 'next/server';
+import { clampPolicy, type RetentionPolicy } from '@/lib/offline-replica/store';
+import { resolveIndexSession, syncAccount } from '@/lib/offline-replica/engine';
+import { gateReplicaRoute, NO_STORE, replicaErrorResponse } from '@/lib/offline-replica/route-gate';
+
+export const runtime = 'nodejs';
+export const dynamic = 'force-dynamic';
+
+export async function POST(request: NextRequest) {
+  const gated = gateReplicaRoute();
+  if (gated) return gated;
+
+  let body: Record<string, unknown> = {};
+  try {
+    const text = await request.text();
+    if (text.trim()) body = JSON.parse(text) as Record<string, unknown>;
+  } catch {
+    return NextResponse.json({ error: 'Malformed JSON body' }, { status: 400 });
+  }
+
+  const rawPolicy = body.policy;
+  const policy: RetentionPolicy | undefined =
+    rawPolicy && typeof rawPolicy === 'object' && !Array.isArray(rawPolicy)
+      ? clampPolicy(rawPolicy as Partial<RetentionPolicy>)
+      : undefined;
+
+  try {
+    const session = await resolveIndexSession(request);
+    const report = await syncAccount(session, {
+      policy,
+      forceResync: body.forceResync === true,
+    });
+    return NextResponse.json({ ok: report.ok, report }, { headers: NO_STORE });
+  } catch (error) {
+    return replicaErrorResponse(error, 'sync');
+  }
+}
diff --git a/components/settings/local-index-settings.tsx b/components/settings/local-index-settings.tsx
index 670dbeac..59e529ce 100644
--- a/components/settings/local-index-settings.tsx
+++ b/components/settings/local-index-settings.tsx
@@ -15,6 +15,10 @@ import { SettingsSection, SettingItem } from './settings-section';
 import { isElectronShell } from '@/lib/electron-bridge';
 import { useAccountStore } from '@/stores/account-store';
 import { catchUpIndex, fetchIndexStats, type IndexStats } from '@/lib/mail-index-client';
+import {
+  chainSync, fetchReplicaStatus, purgeReplica, updateRetentionPolicy,
+  type ReplicaStatus, type RetentionPolicy,
+} from '@/lib/offline-replica-client';
 
 const TYPE_LABELS: Record<string, string> = {
   mail: 'Mail',
@@ -118,6 +122,190 @@ export function LocalIndexSettings() {
           {busy ? 'Indexing…' : 'Update index'}
         </Button>
       </SettingItem>
+
+      <OfflineMailSettings slot={slot} />
     </SettingsSection>
   );
 }
+
+function formatBytes(bytes: number): string {
+  if (bytes < 1024) return `${bytes} B`;
+  const units = ['KB', 'MB', 'GB'];
+  let value = bytes / 1024;
+  let unit = 0;
+  while (value >= 1024 && unit < units.length - 1) { value /= 1024; unit++; }
+  return `${value.toFixed(value >= 10 ? 0 : 1)} ${units[unit]}`;
+}
+
+const PHASE_LABELS: Record<string, string> = {
+  'never-run': 'not started',
+  scanning: 'downloading history',
+  reconciling: 'rebuilding',
+  complete: 'up to date',
+};
+
+/**
+ * Controls for the offline mail replica (lib/offline-replica/**).
+ *
+ * Lives inside the same panel as the search index because they share one
+ * encrypted file, one key and one purge - presenting them as two unrelated
+ * features would misrepresent what "delete" deletes.
+ */
+function OfflineMailSettings({ slot }: { slot: number | undefined }) {
+  const [status, setStatus] = useState<ReplicaStatus | null>(null);
+  const [busy, setBusy] = useState<null | 'sync' | 'purge' | 'policy'>(null);
+  const [message, setMessage] = useState<string | null>(null);
+
+  const refresh = useCallback(async () => {
+    setStatus(await fetchReplicaStatus(slot));
+  }, [slot]);
+
+  useEffect(() => { void refresh(); }, [refresh]);
+
+  const savePolicy = async (patch: Partial<RetentionPolicy>) => {
+    if (!status) return;
+    const next: RetentionPolicy = { ...status.policy, ...patch };
+    setBusy('policy');
+    setMessage(null);
+    try {
+      const ok = await updateRetentionPolicy(next, slot);
+      if (!ok) { setMessage('Could not save the retention setting.'); return; }
+      // The change is applied by the next cycle - a widen re-scans, a narrow
+      // evicts - so run one now rather than leaving the number looking wrong.
+      await chainSync({ slot, max: 2 });
+      await refresh();
+    } finally {
+      setBusy(null);
+    }
+  };
+
+  const handleSync = async () => {
+    setBusy('sync');
+    setMessage(null);
+    try {
+      const report = await chainSync({ slot });
+      if (!report) { setMessage('Offline mail is unavailable on this system.'); return; }
+      setMessage(
+        report.ok
+          ? `Synced ${report.envelopesWritten} messages and ${report.bodiesWritten} bodies.` +
+            (report.unfinishedWork ? ' More will download in the background.' : '') +
+            (report.warnings.length > 0 ? ` Notes: ${report.warnings.join('; ')}` : '')
+          : `Sync failed: ${report.error ?? 'unknown error'}`,
+      );
+      await refresh();
+    } finally {
+      setBusy(null);
+    }
+  };
+
+  const handlePurge = async () => {
+    setBusy('purge');
+    setMessage(null);
+    try {
+      const ok = await purgeReplica(slot);
+      setMessage(ok ? 'Offline mail deleted from this device.' : 'Could not delete offline mail.');
+      await refresh();
+    } finally {
+      setBusy(null);
+    }
+  };
+
+  if (!status) return null;
+
+  const stats = status.stats;
+  const total = stats ? stats.fileBytes : 0;
+
+  return (
+    <>
+      <SettingItem
+        label="Offline mail"
+        description={
+          stats
+            ? `${stats.envelopes} messages listed, ${stats.bodies} readable offline · ` +
+              `${formatBytes(stats.bodyBytes)} of message content · ` +
+              `status: ${PHASE_LABELS[status.coveragePhase] ?? status.coveragePhase}` +
+              (status.resyncRequired ? ' (a rebuild is queued)' : '') +
+              (stats.wantedBodies > 0 ? ` · ${stats.wantedBodies} still downloading` : '')
+            : 'Nothing stored yet. Mail downloads automatically as it arrives.'
+        }
+      >
+        <span className="text-sm text-muted-foreground tabular-nums">{formatBytes(total)}</span>
+      </SettingItem>
+
+      <SettingItem
+        label="Keep message list for"
+        description={
+          'How far back the offline message list goes. Listed messages are tiny (about a ' +
+          'kilobyte each), so a wide window here costs very little and means a message never ' +
+          'disappears from the offline list just because its content was removed to save space.'
+        }
+      >
+        <select
+          className="h-9 rounded-md border border-border bg-background px-2 text-sm"
+          value={status.policy.envelopeDays}
+          disabled={busy !== null}
+          onChange={(e) => void savePolicy({ envelopeDays: Number(e.target.value) })}
+        >
+          {[30, 90, 180, 365, 730, 1825].map((d) => (
+            <option key={d} value={d}>
+              {d >= 365 ? `${Math.round(d / 365)} year${d >= 730 ? 's' : ''}` : `${d} days`}
+            </option>
+          ))}
+        </select>
+      </SettingItem>
+
+      <SettingItem
+        label="Keep full messages for"
+        description={
+          'How far back complete messages - including formatted content - are stored so they ' +
+          'can be read with no network. Attachments are not downloaded; they still need a ' +
+          'connection.'
+        }
+      >
+        <select
+          className="h-9 rounded-md border border-border bg-background px-2 text-sm"
+          value={status.policy.bodyDays}
+          disabled={busy !== null}
+          onChange={(e) => void savePolicy({ bodyDays: Number(e.target.value) })}
+        >
+          {[7, 14, 30, 90, 180, 365].map((d) => (
+            <option key={d} value={d}>{d >= 365 ? '1 year' : `${d} days`}</option>
+          ))}
+        </select>
+      </SettingItem>
+
+      <SettingItem
+        label="Storage limit for message content"
+        description={
+          'The oldest stored content is removed first when this is reached. Messages stay in ' +
+          'the offline list either way - only their content is removed.'
+        }
+      >
+        <select
+          className="h-9 rounded-md border border-border bg-background px-2 text-sm"
+          value={status.policy.maxBodyMB}
+          disabled={busy !== null}
+          onChange={(e) => void savePolicy({ maxBodyMB: Number(e.target.value) })}
+        >
+          {[100, 250, 500, 1000, 2000, 5000].map((mb) => (
+            <option key={mb} value={mb}>{mb >= 1000 ? `${mb / 1000} GB` : `${mb} MB`}</option>
+          ))}
+        </select>
+      </SettingItem>
+
+      <SettingItem
+        label="Offline mail actions"
+        description={message ?? 'Download now, or delete everything stored offline on this device.'}
+      >
+        <div className="flex gap-2">
+          <Button variant="outline" size="sm" onClick={handleSync} disabled={busy !== null}>
+            {busy === 'sync' ? 'Downloading…' : 'Download now'}
+          </Button>
+          <Button variant="outline" size="sm" onClick={handlePurge} disabled={busy !== null}>
+            {busy === 'purge' ? 'Deleting…' : 'Delete offline mail'}
+          </Button>
+        </div>
+      </SettingItem>
+    </>
+  );
+}
diff --git a/docs/ELECTRON-OFFLINE-ENGINE-DESIGN.md b/docs/ELECTRON-OFFLINE-ENGINE-DESIGN.md
index e84b5ece..b832f07a 100644
--- a/docs/ELECTRON-OFFLINE-ENGINE-DESIGN.md
+++ b/docs/ELECTRON-OFFLINE-ENGINE-DESIGN.md
@@ -1,3 +1,37 @@
+> # ⚠️ PARTLY REINSTATED — read this note before the SUPERSEDED banner below
+>
+> A real offline mail replica **now exists**: `lib/offline-replica/**` +
+> `app/api/offline/{sync,mail,status}` + `lib/offline-fallback-client.ts`. So the banner below
+> ("that scope was dropped") is history, not current state. What was reinstated is the DATA MODEL
+> and the SYNC PROTOCOL from this document; what was **not** reinstated is its process
+> architecture — and that distinction is the whole reason the review's critical findings did not
+> come back with it.
+>
+> | This document proposed | What was built |
+> |---|---|
+> | A persistent background worker holding credentials for the process lifetime | **No worker.** A cycle is request-scoped work in an API route using the request's own `jmap_stalwart_ctx` cookie, triggered by the renderer's live push connection — the same shape the search index already uses. |
+> | An epoch-fenced multi-account `registry.json` | **No registry, no epochs.** Single-flight per account inside one process; all state in the SQLite file under a real transaction. |
+> | Multi-account simultaneous sync | **One request, one account, one cycle**, with hard budgets. |
+> | Engine reads a renderer setting to decide whether to sync | **The renderer decides when to sync.** The retention *policy* is durable inside the encrypted store (`/api/offline/status`), because the retention pass genuinely needs it server-side. |
+> | An offline read layer with no coherence story for the webmail's local unread-count arithmetic | **The replica is a FALLBACK, never a cache in front of the server** — consulted only after a read has failed at the transport level, so an online session never sees a replica count. This is the answer to review finding H3, the one finding that genuinely returned. |
+>
+> Consequently: **C2, C3, C4, H1 and H4 remain moot** (they were all consequences of the worker,
+> the registry, or a server-side engine reading renderer state), **C1 and H2 remain fixed** by what
+> already shipped (`optionalDependencies` + guarded require; the key on an inherited fd), and
+> **H3 is now in scope and answered** as above. See `lib/offline-replica/sync.ts`'s header for the
+> finding-by-finding version of this table, kept next to the code it constrains.
+>
+> Two implementation bugs from the mobile client are regressed by name, because both fail silently
+> and both cost user data: the **body-tier infinite redownload loop** (durable `gave_up` +
+> `shed-by-cap` marks + inserted-not-attempted counting) and the **clock-jump guard that wiped the
+> store** (persist the floor that was USED, never the one that was rejected, plus a separate
+> `evictionAllowed` bit). Tests:
+> `lib/offline-replica/__tests__/{store,retention}.test.ts`, and the real network-cut proof in
+> `integration/tests/13-electron-offline-replica.spec.ts`.
+>
+> Still deliberately out of scope: attachment blobs, offline compose/outbox, delegated/shared
+> accounts, and calendar/contacts/files replication.
+
 > # ⚠️ SUPERSEDED — this is not what was built
 >
 > This document designs a **full offline mail replica**: a persistent background sync engine with
diff --git a/docs/ELECTRON-OFFLINE-ENGINE-REVIEW.md b/docs/ELECTRON-OFFLINE-ENGINE-REVIEW.md
index e5460bb6..f5bb4915 100644
--- a/docs/ELECTRON-OFFLINE-ENGINE-REVIEW.md
+++ b/docs/ELECTRON-OFFLINE-ENGINE-REVIEW.md
@@ -1,3 +1,31 @@
+> # ⚠️ UPDATE — a replica was later built, and this review is why it is shaped the way it is
+>
+> The table below says most of these findings "stopped existing" because the scope change removed
+> the thing they were about. A replica has since been built (`lib/offline-replica/**`), so that
+> reasoning was re-examined finding by finding rather than inherited:
+>
+> - **C1** — still FIXED, and untouched: the replica adds no new dependency and reuses the guarded
+>   optional require. Both `docker build`s are unaffected.
+> - **C2, C3, C4, H1, H4** — still MOOT, and moot *for the same reasons*, because the persistent
+>   background worker, the shared registry and the server-side-engine-reads-renderer-state shapes
+>   were **not** reinstated. A cycle is request-scoped work in an API route with no resident
+>   credential; there is no registry and no epoch; one request syncs one account. Had the worker
+>   come back, all five would have come back with it.
+> - **H2** — still FIXED: the key crosses on an inherited file descriptor, never via environment,
+>   and is zeroed after each job. The replica reuses that channel rather than inventing a second.
+> - **H3 — BACK IN SCOPE, and the only one that is.** This review was right that the webmail does
+>   local delta arithmetic on mailbox unread counts, and a read-only offline cache underneath it
+>   needs a coherence story. The answer is an ordering rule: the replica is consulted **only after
+>   a read has failed at the transport level**, so it is never a cache in front of the server and
+>   the arithmetic never operates on replica numbers. Enforcing that needed a real signal, because
+>   `lib/jmap/client.ts` swallows read errors and returns plausible success — hence
+>   `lib/jmap/transport-health.ts` and the two-part gate in `lib/offline-fallback-client.ts`.
+> - The *medium/low* findings (Linux-only API, the vacuous `cipher_version` check, the two bindings
+>   not being interchangeable) were all already fixed in the shipped index and are inherited.
+>
+> Nothing in this review turned out to be wrong on re-reading. Its verdict — that the sync-engine
+> core transfers and the platform-specific sections were where the danger lay — held exactly.
+
 > # ⚠️ SUPERSEDED — reviews a design that was not built
 >
 > This reviews `ELECTRON-OFFLINE-ENGINE-DESIGN.md`, which was **dropped**. Its findings were the
diff --git a/integration/tests/13-electron-offline-replica.spec.ts b/integration/tests/13-electron-offline-replica.spec.ts
new file mode 100644
index 00000000..aa66ee0c
--- /dev/null
+++ b/integration/tests/13-electron-offline-replica.spec.ts
@@ -0,0 +1,600 @@
+import { test, expect, _electron as electron } from '@playwright/test';
+import type { ElectronApplication, Page } from '@playwright/test';
+import { spawn, type ChildProcess } from 'node:child_process';
+import { createServer } from 'node:net';
+import { get as httpGet } from 'node:http';
+import { createHash, randomBytes } from 'node:crypto';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { ACCOUNTS, JMAP_URL } from './helpers/config';
+import { sendMail } from './helpers/smtp';
+import { JmapClient } from './helpers/jmap';
+
+/**
+ * The offline mail replica (lib/offline-replica/**) against the real Stalwart
+ * fixture, with a REAL NETWORK CUT.
+ *
+ * THE POINT OF THIS FILE: a sync test that never tests the offline case has not
+ * tested the feature. So test 1 syncs against a live server, then makes the
+ * backend genuinely unreachable, and only then asserts that a previously-synced
+ * message still returns its full HTML body - from the encrypted replica, with no
+ * network available to fall back to.
+ *
+ * HOW THE CUT IS MADE. The standalone server is started with
+ * `JMAP_SERVER_URL` pointing at a LOCAL PROXY that forwards to Stalwart. Killing
+ * the proxy's listener makes every JMAP request fail with ECONNREFUSED - a real
+ * transport failure at the socket level, not a mock, not a stubbed fetch, and not
+ * a flag the code under test can see. Preferred over stopping the Stalwart
+ * container because it cuts only THIS test's path and leaves the shared fixture
+ * (and any concurrently-running suite) untouched.
+ *
+ * THE SAME TWO CONSTRAINTS as 12-electron-mail-index.spec.ts apply and are why
+ * this is split into two tests rather than one:
+ *
+ *  1. The RENDERER cannot reach this fixture from a production build. It talks
+ *     JMAP directly to Stalwart, which here is deliberately plain HTTP, and the
+ *     production CSP pins `connect-src` to `'self' https: wss:`. NODE_ENV at
+ *     runtime does not help - `next build` inlines it into the middleware.
+ *  2. The fd-3 key channel cannot survive `next dev`, which claims fd 3 for its
+ *     own IPC. So the two configurations are mutually exclusive: a real key
+ *     channel means no browser, a browser means no key channel.
+ *
+ * Test 1 therefore drives the REAL standalone server over HTTP from Node with a
+ * real fd-3 key channel - no browser needed, because the routes are the thing
+ * being proven. Test 2 launches the REAL Electron shell to prove the routes exist
+ * and are reachable in a genuine build, which is the class of failure only a real
+ * build reveals (the standalone output silently dropping a native prebuild, say).
+ *
+ * WHAT THIS FILE DOES NOT PROVE: that `components/email/email-viewer.tsx` paints
+ * the replica-served body in a browser while offline. That needs a renderer, a
+ * key channel and a reachable-then-unreachable JMAP server simultaneously, which
+ * constraints 1 and 2 make impossible against this fixture. The read path returns
+ * a field-for-field `Email` (asserted below, including `bodyValues` keyed by the
+ * same partIds as `htmlBody`), and the renderer-side gate is covered by
+ * `lib/__tests__/offline-fallback-client.test.ts` - but the final paint is NOT
+ * covered by a real offline browser run. Stated rather than implied.
+ */
+const alice = ACCOUNTS.alice;
+const projectRoot = path.resolve(__dirname, '../..');
+
+/** Mirrors lib/mail-index/paths.ts's accountFileToken(). */
+function accountFileToken(accountId: string): string {
+  return createHash('sha256').update(accountId, 'utf8').digest('hex').slice(0, 32);
+}
+
+function getFreePort(): Promise<number> {
+  return new Promise((resolve, reject) => {
+    const server = createServer();
+    server.unref();
+    server.on('error', reject);
+    server.listen(0, '127.0.0.1', () => {
+      const address = server.address();
+      if (address && typeof address === 'object') {
+        const { port } = address;
+        server.close(() => resolve(port));
+      } else {
+        server.close(() => reject(new Error('Could not allocate a free localhost port')));
+      }
+    });
+  });
+}
+
+function waitForServerReady(url: string, timeoutMs: number): Promise<void> {
+  const deadline = Date.now() + timeoutMs;
+  return new Promise((resolve, reject) => {
+    const attempt = () => {
+      const req = httpGet(url, (res) => {
+        res.resume();
+        resolve();
+      });
+      req.on('error', () => {
+        if (Date.now() > deadline) {
+          reject(new Error(`Server never became reachable at ${url}`));
+          return;
+        }
+        setTimeout(attempt, 300);
+      });
+    };
+    attempt();
+  });
+}
+
+/**
+ * A raw TCP forwarder in front of Stalwart, so the test can sever the backend at
+ * the socket level. `cut()` closes the listener AND destroys every live socket, so
+ * a pooled keep-alive connection cannot keep working after the cut.
+ */
+async function startCuttableProxy(target: { host: string; port: number }): Promise<{
+  port: number;
+  cut: () => Promise<void>;
+  stop: () => Promise<void>;
+}> {
+  const { connect } = await import('node:net');
+  const sockets = new Set<import('node:net').Socket>();
+  const server = createServer((incoming) => {
+    sockets.add(incoming);
+    incoming.on('close', () => sockets.delete(incoming));
+    incoming.on('error', () => incoming.destroy());
+    const upstream = connect(target.port, target.host, () => {
+      incoming.pipe(upstream);
+      upstream.pipe(incoming);
+    });
+    sockets.add(upstream);
+    upstream.on('close', () => sockets.delete(upstream));
+    upstream.on('error', () => { incoming.destroy(); upstream.destroy(); });
+  });
+  const port = await getFreePort();
+  await new Promise<void>((resolve, reject) => {
+    server.once('error', reject);
+    server.listen(port, '127.0.0.1', () => resolve());
+  });
+
+  const closeAll = () =>
+    new Promise<void>((resolve) => {
+      for (const s of sockets) s.destroy();
+      sockets.clear();
+      server.close(() => resolve());
+      // `close()` only stops new connections; the destroys above handle the rest.
+      setTimeout(resolve, 500);
+    });
+
+  return { port, cut: closeAll, stop: closeAll };
+}
+
+/**
+ * Serves the key protocol of electron/key-service.ts over the child's inherited
+ * fd. The key and the encryption are real; only safeStorage's wrapping of it is
+ * out of the picture here, which is what test 2 covers.
+ */
+function serveKeyChannel(child: ChildProcess, fdIndex: number, key: Buffer): void {
+  const channel = child.stdio[fdIndex] as NodeJS.ReadWriteStream | null;
+  if (!channel) throw new Error(`child has no fd ${fdIndex} - key channel missing`);
+  let buffer = '';
+  channel.on('data', (chunk: Buffer) => {
+    buffer += chunk.toString('utf8');
+    let newline: number;
+    while ((newline = buffer.indexOf('\n')) >= 0) {
+      const line = buffer.slice(0, newline);
+      buffer = buffer.slice(newline + 1);
+      if (!line.trim()) continue;
+      const req = JSON.parse(line) as { id?: number; op?: string };
+      const reply =
+        req.op === 'getIndexKey'
+          ? { id: req.id, ok: true, key: key.toString('hex') }
+          : req.op === 'deleteIndexKey'
+            ? { id: req.id, ok: true }
+            : { id: req.id, ok: false, code: 'bad-request', error: 'unknown op' };
+      channel.write(`${JSON.stringify(reply)}\n`);
+    }
+  });
+}
+
+class Jar {
+  private cookies = new Map<string, string>();
+  absorb(response: Response): void {
+    for (const raw of response.headers.getSetCookie()) {
+      const [pair] = raw.split(';');
+      const eq = pair.indexOf('=');
+      if (eq <= 0) continue;
+      this.cookies.set(pair.slice(0, eq).trim(), pair.slice(eq + 1).trim());
+    }
+  }
+  header(): string {
+    return [...this.cookies].map(([k, v]) => `${k}=${v}`).join('; ');
+  }
+}
+
+interface CycleReport {
+  ok: boolean;
+  unfinishedWork: boolean;
+  bootstrapped: boolean;
+  envelopesWritten: number;
+  bodiesWritten: number;
+  envelopesDeleted: number;
+  coveragePhase: string;
+  resyncRequired: boolean;
+  warnings: string[];
+  error?: string;
+  errorClass?: string;
+}
+
+test.describe('Electron desktop shell - offline mail replica', () => {
+  test('syncs full bodies, then serves a synced message with the backend UNREACHABLE', async () => {
+    test.setTimeout(240_000);
+
+    const jmap = await JmapClient.connect(alice.email, alice.password);
+    await jmap.reset();
+
+    const stamp = Date.now();
+    const subject = `IT replica subject ${stamp}`;
+    // Appears ONLY in the HTML body, so a hit proves the full body was stored -
+    // not the preview or the subject, which any envelope already carries.
+    const bodyPhrase = `luzernrenewal${stamp}`;
+    const htmlMarker = `<strong>${bodyPhrase}</strong>`;
+
+    await sendMail({
+      from: alice.email,
+      authPass: alice.password,
+      to: alice.email,
+      subject,
+      body: `plain text ${bodyPhrase}`,
+      html: `<html><body><p>Please review the ${htmlMarker} before September.</p></body></html>`,
+    });
+    // A second message, so "the list came from the replica" is not a one-row
+    // coincidence.
+    await sendMail({
+      from: alice.email,
+      authPass: alice.password,
+      to: alice.email,
+      subject: `IT replica second ${stamp}`,
+      body: 'the second message',
+    });
+
+    const storeDir = fs.mkdtempSync(path.join(os.tmpdir(), 'vncmail-replica-it-'));
+    const key = randomBytes(32);
+    const stalwart = new URL(JMAP_URL);
+    const proxy = await startCuttableProxy({
+      host: stalwart.hostname,
+      port: Number(stalwart.port || 80),
+    });
+    const proxiedJmapUrl = `http://127.0.0.1:${proxy.port}`;
+
+    const port = await getFreePort();
+    const baseUrl = `http://127.0.0.1:${port}`;
+    const serverEntry = path.join(projectRoot, '.next', 'standalone', 'server.js');
+    expect(
+      fs.existsSync(serverEntry),
+      `missing ${serverEntry} - run "npm run build:standalone" first`,
+    ).toBe(true);
+
+    const server = spawn(process.execPath, [serverEntry], {
+      cwd: path.dirname(serverEntry),
+      env: {
+        ...process.env,
+        PORT: String(port),
+        HOSTNAME: '127.0.0.1',
+        NODE_ENV: 'production',
+        // Through the cuttable proxy, so the backend can be severed later.
+        JMAP_SERVER_URL: proxiedJmapUrl,
+        SESSION_SECRET: 'integration-not-a-real-secret-32-chars!!',
+        VNCMAIL_DESKTOP_STORE_DIR: storeDir,
+        VNCMAIL_DESKTOP_KEY_FD: '3',
+      },
+      stdio: ['pipe', 'pipe', 'pipe', 'pipe'],
+    });
+    server.stderr?.on('data', (chunk) => process.stderr.write(`[standalone] ${chunk}`));
+    serveKeyChannel(server, 3, key);
+
+    const jar = new Jar();
+    const call = async (url: string, init?: RequestInit): Promise<Response> => {
+      const response = await fetch(`${baseUrl}${url}`, {
+        ...init,
+        headers: { ...(init?.headers ?? {}), cookie: jar.header() },
+      });
+      jar.absorb(response);
+      return response;
+    };
+    const sync = async (body: Record<string, unknown> = {}): Promise<CycleReport> => {
+      const response = await call('/api/offline/sync', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(body),
+      });
+      const parsed = await response.json();
+      expect(response.status, JSON.stringify(parsed)).toBe(200);
+      return parsed.report as CycleReport;
+    };
+
+    try {
+      await waitForServerReady(baseUrl, 90_000);
+
+      const login = await call('/api/auth/session?slot=0', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({
+          serverUrl: proxiedJmapUrl,
+          username: alice.email,
+          password: alice.password,
+          slot: 0,
+        }),
+      });
+      const loginText = await login.text();
+      expect(login.status, `login failed: ${loginText}`).toBe(200);
+
+      // The gate must be open and the native binding loaded, or every assertion
+      // below would fail for an unrelated reason.
+      const reachable = await call('/api/offline/status');
+      const reachableText = await reachable.text();
+      expect(
+        reachable.status,
+        `replica routes unreachable: ${reachableText.slice(0, 400)}`,
+      ).toBe(200);
+
+      // ── ONLINE: bootstrap, then chain until the cycle reports itself done ──
+      const first = await sync();
+      expect(first.error, `first cycle failed: ${first.error}`).toBeUndefined();
+      expect(first.bootstrapped, 'the first cycle must bootstrap').toBe(true);
+
+      let report = first;
+      for (let i = 0; i < 12 && report.unfinishedWork; i++) report = await sync();
+      expect(
+        report.unfinishedWork,
+        `sync never settled: ${JSON.stringify(report)}`,
+      ).toBe(false);
+      // Termination is a real property here: the body-queue give-up marks and the
+      // inserted-not-attempted count are what stop this looping forever.
+      expect(report.coveragePhase).toBe('complete');
+      expect(report.resyncRequired).toBe(false);
+
+      const status = await (await call('/api/offline/status')).json();
+      expect(status.synced).toBe(true);
+      expect(
+        status.stats.envelopes,
+        `no envelopes stored: ${JSON.stringify(status.stats)}`,
+      ).toBeGreaterThanOrEqual(2);
+      expect(
+        status.stats.bodies,
+        `no BODIES stored - the replica would be no better than the search index`,
+      ).toBeGreaterThanOrEqual(2);
+      expect(status.stats.mailboxes).toBeGreaterThan(0);
+
+      // ── THE DELTA PATH: a message that arrives AFTER the cursor was captured ──
+      // Bootstrap alone would satisfy every assertion below, so this is what actually
+      // exercises `Email/changes` and proves the stored cursor is USABLE rather than
+      // merely present. It is also the assertion that fails if an `Email/get` state
+      // token is ever adopted as a `/changes` cursor: the fast-forwarded cursor
+      // reports no changes, and this message never arrives.
+      const deltaSubject = `IT replica delta ${stamp}`;
+      const deltaPhrase = `bernrenewal${stamp}`;
+      await sendMail({
+        from: alice.email,
+        authPass: alice.password,
+        to: alice.email,
+        subject: deltaSubject,
+        body: `plain ${deltaPhrase}`,
+        html: `<html><body><p>delta ${deltaPhrase}</p></body></html>`,
+      });
+
+      let delta = await sync();
+      for (let i = 0; i < 10 && (delta.unfinishedWork || delta.envelopesWritten === 0); i++) {
+        delta = await sync();
+      }
+      expect(delta.bootstrapped, 'the delta cycle must NOT re-bootstrap').toBe(false);
+      const afterDelta = await (await call('/api/offline/status')).json();
+      expect(
+        afterDelta.stats.envelopes,
+        `Email/changes did not deliver a message that arrived after the cursor was ` +
+          `captured: ${JSON.stringify(afterDelta.stats)}`,
+      ).toBeGreaterThanOrEqual(3);
+      expect(
+        afterDelta.stats.bodies,
+        'the delta path delivered the envelope but never queued its body',
+      ).toBeGreaterThanOrEqual(3);
+      expect(afterDelta.resyncRequired, 'a healthy delta cycle must not invalidate a cursor').toBe(false);
+
+      // Find the message and its mailbox while still online, so the offline phase
+      // asserts on known ids rather than discovering them from the thing under test.
+      const mailboxesOnline = await (await call('/api/offline/mail?kind=mailboxes')).json();
+      const inbox = (mailboxesOnline.mailboxes as Array<{ id: string; role?: string }>)
+        .find((m) => m.role === 'inbox');
+      expect(inbox, 'the replica holds no inbox').toBeTruthy();
+
+      const listOnline = await (
+        await call(`/api/offline/mail?kind=list&mailboxId=${encodeURIComponent(inbox!.id)}&limit=50`)
+      ).json();
+      const target = (listOnline.emails as Array<{ id: string; subject?: string }>)
+        .find((e) => e.subject === subject);
+      expect(target, `the synced message is not in the replica: ${JSON.stringify(listOnline.emails?.map((e: {subject?: string}) => e.subject))}`).toBeTruthy();
+
+      // ── THE CUT: sever the backend at the socket level ────────────────────
+      await proxy.cut();
+
+      // Prove the cut is real, from inside the server process's own network
+      // namespace: a live JMAP call must now fail. `/api/offline/sync` reaches
+      // Stalwart first thing, so it is the honest probe.
+      const afterCut = await call('/api/offline/sync', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({}),
+      });
+      const afterCutBody = await afterCut.json();
+      expect(
+        afterCut.status,
+        `the backend is still reachable, so the offline assertions below would prove nothing: ` +
+          `${JSON.stringify(afterCutBody)}`,
+      ).not.toBe(200);
+
+      // ── OFFLINE: the actual feature ───────────────────────────────────────
+      const messageResponse = await call(
+        `/api/offline/mail?kind=message&id=${encodeURIComponent(target!.id)}`,
+      );
+      // Read the body ONCE: `expect`'s message argument is evaluated eagerly, so
+      // putting `await response.text()` in it consumes the stream before .json().
+      const messageText = await messageResponse.text();
+      expect(
+        messageResponse.status,
+        `the offline read path failed with the backend down: ${messageText.slice(0, 400)}`,
+      ).toBe(200);
+      const offline = JSON.parse(messageText);
+      expect(offline.available).toBe(true);
+      expect(offline.hasBody, 'the message has no stored body offline').toBe(true);
+
+      const email = offline.email as {
+        id: string; subject?: string; receivedAt: string;
+        htmlBody?: Array<{ partId: string; type: string }>;
+        textBody?: Array<{ partId: string }>;
+        bodyValues?: Record<string, { value: string }>;
+        from?: Array<{ email: string }>;
+        keywords?: Record<string, boolean>;
+        mailboxIds?: Record<string, boolean>;
+        headers?: Record<string, string | string[]>;
+      };
+
+      expect(email.id).toBe(target!.id);
+      expect(email.subject).toBe(subject);
+
+      // THE ASSERTION: the full HTML body, recovered with no network.
+      const htmlPartId = email.htmlBody?.[0]?.partId;
+      expect(htmlPartId, 'no htmlBody part offline').toBeTruthy();
+      const html = email.bodyValues?.[htmlPartId as string]?.value ?? '';
+      expect(
+        html,
+        'the HTML body is not in the replica - this is the whole feature',
+      ).toContain(htmlMarker);
+      expect(html).toContain(bodyPhrase);
+
+      // `bodyValues` MUST be keyed by the same partIds as htmlBody/textBody, or
+      // email-viewer.tsx's isBodyLoading gate sits on its skeleton forever
+      // (hasBodyParts true, bodyValues unusable).
+      for (const part of [...(email.htmlBody ?? []), ...(email.textBody ?? [])]) {
+        expect(
+          email.bodyValues?.[part.partId],
+          `bodyValues is missing partId ${part.partId}, which the viewer requires`,
+        ).toBeTruthy();
+      }
+
+      // The rest of the shape the renderer reads.
+      expect(email.from?.[0]?.email).toBe(alice.email);
+      expect(email.receivedAt).toBeTruthy();
+      expect(Object.keys(email.mailboxIds ?? {})).toContain(inbox!.id);
+      // Header normalisation happened server-side (the array -> record flattening
+      // the online path does in parseEmailHeaders).
+      expect(email.headers && !Array.isArray(email.headers)).toBe(true);
+
+      // The list and the folder tree must also survive the cut.
+      const listOffline = await (
+        await call(`/api/offline/mail?kind=list&mailboxId=${encodeURIComponent(inbox!.id)}&limit=50`)
+      ).json();
+      expect(listOffline.available).toBe(true);
+      expect(listOffline.emails.length).toBeGreaterThanOrEqual(2);
+      expect(
+        (listOffline.emails as Array<{ subject?: string }>).map((e) => e.subject),
+      ).toContain(subject);
+
+      const mailboxesOffline = await (await call('/api/offline/mail?kind=mailboxes')).json();
+      expect(mailboxesOffline.available).toBe(true);
+      expect((mailboxesOffline.mailboxes as unknown[]).length).toBeGreaterThan(0);
+
+      // Status must be readable offline too - a user with no network still needs
+      // to see what they have and be able to free the space.
+      const statusOffline = await (await call('/api/offline/status')).json();
+      expect(statusOffline.ok).toBe(true);
+      expect(statusOffline.stats.bodies).toBeGreaterThanOrEqual(2);
+
+      // A cycle attempted while offline must classify as Transport and must NOT
+      // touch the data. "Offline is not an error."
+      expect(
+        ['Transport', 'ServerTransient'].includes(String(afterCutBody.code)),
+        `an offline cycle must classify as Transport/ServerTransient so the caller retries ` +
+          `rather than treating the feature as broken; got code=${afterCutBody.code} ` +
+          `status=${afterCut.status} body=${JSON.stringify(afterCutBody)}`,
+      ).toBe(true);
+      const afterOfflineCycle = await (await call('/api/offline/status')).json();
+      expect(
+        afterOfflineCycle.stats.envelopes,
+        'an offline cycle deleted data - a transport failure must never do that',
+      ).toBe(statusOffline.stats.envelopes);
+      expect(afterOfflineCycle.resyncRequired).toBe(false);
+
+      // ── PURGE: the retention control has to actually free the space ────────
+      const purge = await call('/api/offline/status', { method: 'DELETE' });
+      expect(purge.status).toBe(200);
+      const purged = await (await call('/api/offline/status')).json();
+      expect(purged.synced).toBe(false);
+      expect(purged.coveragePhase).toBe('never-run');
+    } finally {
+      server.kill();
+      await proxy.stop();
+      // Let the process release its WAL files before reading them.
+      await new Promise((r) => setTimeout(r, 700));
+    }
+
+    // ── the file on disk is genuinely encrypted ─────────────────────────────
+    const accountId = `${alice.email}@127.0.0.1`;
+    const dbPath = path.join(storeDir, 'index', `${accountFileToken(accountId)}.db`);
+    expect(fs.existsSync(dbPath), `no replica database at ${dbPath}`).toBe(true);
+
+    const onDisk = Buffer.concat(
+      ['', '-wal', '-shm']
+        .map((suffix) => `${dbPath}${suffix}`)
+        .filter((f) => fs.existsSync(f))
+        .map((f) => fs.readFileSync(f)),
+    );
+    expect(onDisk.length).toBeGreaterThan(0);
+    // `PRAGMA key` is a silent no-op on a non-SQLCipher binding - no error, a
+    // working database, and the mail in cleartext - so every functional assertion
+    // above would pass either way. These are the ones that catch it.
+    expect(
+      fs.readFileSync(dbPath).subarray(0, 15).toString('latin1'),
+      'the replica file has a plain SQLite header - it is NOT encrypted',
+    ).not.toBe('SQLite format 3');
+    expect(
+      onDisk.includes(bodyPhrase),
+      'the message body is recoverable from the raw database bytes - not encrypted',
+    ).toBe(false);
+    expect(
+      onDisk.includes(subject),
+      'the subject is recoverable from the raw database bytes - not encrypted',
+    ).toBe(false);
+
+    fs.rmSync(storeDir, { recursive: true, force: true });
+  });
+
+  test('wiring: the real standalone boot reaches the replica routes with a real safeStorage key', async () => {
+    test.setTimeout(180_000);
+    // A FRESH profile is load-bearing, not hygiene: the 401 asserted below is
+    // "nobody is signed in", and a leftover jmap_stalwart_ctx cookie from any
+    // previous run turns it into a 200.
+    const userDataDir = fs.mkdtempSync(path.join(os.tmpdir(), 'vncmail-replica-wiring-'));
+    const electronApp: ElectronApplication = await electron.launch({
+      args: [projectRoot, `--user-data-dir=${userDataDir}`],
+      env: {
+        ...process.env,
+        JMAP_SERVER_URL: JMAP_URL,
+        SESSION_SECRET: 'integration-not-a-real-secret-32-chars!!',
+      },
+    });
+
+    try {
+      const appWindow: Page = await electronApp.firstWindow();
+      await appWindow.waitForLoadState('domcontentloaded');
+      await appWindow.locator('#username').waitFor({ state: 'visible', timeout: 90_000 });
+
+      const encryptionAvailable = await electronApp.evaluate(({ safeStorage }) =>
+        safeStorage.isEncryptionAvailable(),
+      );
+      expect(
+        encryptionAvailable,
+        'safeStorage reports no encryption available, so main.ts correctly disabled the ' +
+          'feature - this assertion cannot pass here',
+      ).toBe(true);
+
+      // 401 = the gate opened, the native binding loaded from the REAL standalone
+      // artifact, and the fd-3 key channel is present; it refuses only because
+      // nobody is signed in.
+      //   404 => VNCMAIL_DESKTOP_STORE_DIR was never set
+      //   503 => the native binding or the key channel is missing from the real
+      //          build - the class of failure only a real build reveals
+      for (const route of [
+        '/api/offline/status',
+        '/api/offline/mail?kind=mailboxes',
+        '/api/offline/sync',
+      ]) {
+        const probe = await appWindow.evaluate(async (url) => {
+          const response = await fetch(url, {
+            method: url.endsWith('/sync') ? 'POST' : 'GET',
+          });
+          return { status: response.status, body: (await response.text()).slice(0, 300) };
+        }, route);
+        expect(
+          probe.status,
+          `${route}: expected 401 (reachable, unauthenticated) but got ${probe.status}: ${probe.body}`,
+        ).toBe(401);
+      }
+    } finally {
+      await electronApp.close();
+      fs.rmSync(userDataDir, { recursive: true, force: true });
+    }
+  });
+});
diff --git a/integration/tests/helpers/smtp.ts b/integration/tests/helpers/smtp.ts
index 734c2e00..e7b61387 100644
--- a/integration/tests/helpers/smtp.ts
+++ b/integration/tests/helpers/smtp.ts
@@ -22,6 +22,14 @@ interface SendOptions {
   subject: string;
   /** Plain-text body. */
   body: string;
+  /**
+   * Optional HTML alternative, sent as multipart/alternative alongside `body`.
+   *
+   * Added for 13-electron-offline-replica.spec.ts, which has to prove the offline
+   * replica stores a real HTML body and not just the plain-text excerpt the search
+   * index keeps - so the message needs a genuine distinct text/html part.
+   */
+  html?: string;
   /** Extra headers (e.g. custom Message-ID / In-Reply-To for threading). */
   headers?: Record<string, string>;
   /** Optional single attachment (sent as multipart/mixed, base64). */
@@ -158,6 +166,22 @@ export async function sendMail(opts: SendOptions): Promise<void> {
         b64,
         `--${boundary}--`,
       ].join('\r\n');
+    } else if (opts.html) {
+      const boundary = 'italt_boundary_0001';
+      headers['MIME-Version'] = '1.0';
+      headers['Content-Type'] = `multipart/alternative; boundary="${boundary}"`;
+      // text first, html second: multipart/alternative is least-to-most preferred.
+      mime = [
+        `--${boundary}`,
+        'Content-Type: text/plain; charset=utf-8',
+        '',
+        crlf(opts.body),
+        `--${boundary}`,
+        'Content-Type: text/html; charset=utf-8',
+        '',
+        crlf(opts.html),
+        `--${boundary}--`,
+      ].join('\r\n');
     } else {
       headers['Content-Type'] = 'text/plain; charset=utf-8';
       mime = crlf(opts.body);
diff --git a/lib/__tests__/offline-fallback-client.test.ts b/lib/__tests__/offline-fallback-client.test.ts
new file mode 100644
index 00000000..cc5c6eb4
--- /dev/null
+++ b/lib/__tests__/offline-fallback-client.test.ts
@@ -0,0 +1,224 @@
+// The two-part fallback gate.
+//
+// `lib/jmap/client.ts`'s read methods swallow their own errors and return
+// plausible success, so a "looks empty" result is NOT evidence of a network
+// failure - it is also what a genuinely empty folder returns, and
+// `getMailboxes()` fabricates a synthetic Inbox rather than throwing. Falling back
+// on the shape alone would serve stale replica rows over a folder the user had
+// just emptied. So the gate is: suspicious result AND a `fetch` rejection recorded
+// during that exact call.
+
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+import type { IJMAPClient } from '@/lib/jmap/client-interface';
+import type { Email, Mailbox } from '@/lib/jmap/types';
+import { noteTransportFailure, resetTransportHealth } from '@/lib/jmap/transport-health';
+
+const readOfflineMailboxes = vi.fn();
+const readOfflineList = vi.fn();
+const readOfflineMessage = vi.fn();
+const isReplicaUnavailable = vi.fn(() => false);
+
+vi.mock('@/lib/offline-replica-client', () => ({
+  readOfflineMailboxes: (...a: unknown[]) => readOfflineMailboxes(...a),
+  readOfflineList: (...a: unknown[]) => readOfflineList(...a),
+  readOfflineMessage: (...a: unknown[]) => readOfflineMessage(...a),
+  isReplicaUnavailable: () => isReplicaUnavailable(),
+}));
+
+vi.mock('@/stores/account-store', () => ({
+  useAccountStore: {
+    getState: () => ({
+      accounts: [{ id: 'alice@mail.example.org', cookieSlot: 3, serverIdentifiers: [] }],
+    }),
+  },
+}));
+
+const { withOfflineFallback } = await import('@/lib/offline-fallback-client');
+
+function replicaEmail(id: string): Email {
+  return {
+    id, threadId: 't', mailboxIds: { inbox: true }, keywords: {}, size: 1,
+    receivedAt: '2026-08-01T00:00:00.000Z', hasAttachment: false,
+    htmlBody: [{ partId: '1', blobId: 'b', size: 1, type: 'text/html' }],
+    bodyValues: { '1': { value: '<p>from the replica</p>' } },
+  };
+}
+
+interface Stub extends Partial<IJMAPClient> {
+  getEmail: IJMAPClient['getEmail'];
+  getEmails: IJMAPClient['getEmails'];
+  getMailboxes: IJMAPClient['getMailboxes'];
+  getAllMailboxes: IJMAPClient['getAllMailboxes'];
+}
+
+/** Reproduces the client's real error-swallowing shapes. */
+function stubClient(overrides: Partial<Stub> = {}): IJMAPClient {
+  const stub = {
+    getUsername: () => 'alice',
+    getServerUrl: () => 'https://mail.example.org',
+    getAccountId: () => 'primary',
+    getEmail: async () => null,
+    getEmails: async () => ({ emails: [] as Email[], hasMore: false, total: 0 }),
+    getMailboxes: async () => ([
+      // The exact placeholder client.ts fabricates on failure.
+      { id: 'INBOX', name: 'Inbox', role: 'inbox', sortOrder: 0, totalEmails: 0,
+        unreadEmails: 0, totalThreads: 0, unreadThreads: 0, isSubscribed: true,
+        myRights: {} } as unknown as Mailbox,
+    ]),
+    getAllMailboxes: async () => ([
+      { id: 'INBOX', name: 'Inbox', role: 'inbox', sortOrder: 0, totalEmails: 0,
+        unreadEmails: 0, totalThreads: 0, unreadThreads: 0, isSubscribed: true,
+        myRights: {} } as unknown as Mailbox,
+    ]),
+    ...overrides,
+  };
+  return stub as unknown as IJMAPClient;
+}
+
+describe('withOfflineFallback', () => {
+  beforeEach(() => {
+    resetTransportHealth();
+    vi.clearAllMocks();
+    isReplicaUnavailable.mockReturnValue(false);
+  });
+
+  it('does NOT consult the replica when the server answered "empty"', async () => {
+    // The whole point. An empty folder must render empty, not as whatever the
+    // replica last held.
+    const client = withOfflineFallback(stubClient());
+    const result = await client.getEmails('inbox');
+    expect(result.emails).toEqual([]);
+    expect(readOfflineList).not.toHaveBeenCalled();
+
+    expect(await client.getEmail('e1')).toBeNull();
+    expect(readOfflineMessage).not.toHaveBeenCalled();
+  });
+
+  it('consults the replica when a transport failure happened DURING the call', async () => {
+    readOfflineList.mockResolvedValue({
+      emails: [replicaEmail('e1')], total: 1, hasMore: false,
+    });
+    const client = withOfflineFallback(
+      stubClient({
+        getEmails: async () => {
+          // What authenticatedFetch does when `fetch` rejects.
+          noteTransportFailure();
+          return { emails: [], hasMore: false, total: 0 };
+        },
+      }),
+    );
+    const result = await client.getEmails('inbox', undefined, 25, 0);
+    expect(result.emails.map((e) => e.id)).toEqual(['e1']);
+    expect(result.total).toBe(1);
+    // And it asks for the right slot, so a multi-account shell reads the right file.
+    expect(readOfflineList).toHaveBeenCalledWith('inbox', { limit: 25, offset: 0, slot: 3 });
+  });
+
+  it('ignores a stale transport failure from BEFORE the call', async () => {
+    // The counter is sampled per call precisely so an old failure cannot make a
+    // later successful-but-empty read look offline.
+    noteTransportFailure();
+    const client = withOfflineFallback(stubClient());
+    await client.getEmails('inbox');
+    expect(readOfflineList).not.toHaveBeenCalled();
+  });
+
+  it('serves a full message from the replica, but refuses an envelope-only hit', async () => {
+    // An envelope with no bodyValues would render blank AND leave the viewer's
+    // isBodyLoading gate stuck on its skeleton, which is worse than saying the
+    // message is unavailable.
+    readOfflineMessage.mockResolvedValue({ email: replicaEmail('e1'), hasBody: true });
+    const client = withOfflineFallback(
+      stubClient({ getEmail: async () => { noteTransportFailure(); return null; } }),
+    );
+    const email = await client.getEmail('e1');
+    expect(email?.bodyValues?.['1'].value).toContain('from the replica');
+
+    resetTransportHealth();
+    readOfflineMessage.mockResolvedValue({ email: replicaEmail('e2'), hasBody: false });
+    const client2 = withOfflineFallback(
+      stubClient({ getEmail: async () => { noteTransportFailure(); return null; } }),
+    );
+    expect(await client2.getEmail('e2')).toBeNull();
+  });
+
+  it('recognises the synthetic Inbox placeholder and replaces it', async () => {
+    readOfflineMailboxes.mockResolvedValue([
+      { id: 'mb1', name: 'Inbox', role: 'inbox', sortOrder: 0, totalEmails: 9, unreadEmails: 2,
+        totalThreads: 9, unreadThreads: 2, isSubscribed: true, myRights: {} } as unknown as Mailbox,
+    ]);
+    const client = withOfflineFallback(
+      stubClient({
+        getAllMailboxes: async () => {
+          noteTransportFailure();
+          return [
+            { id: 'INBOX', name: 'Inbox', role: 'inbox', sortOrder: 0, totalEmails: 0,
+              unreadEmails: 0, totalThreads: 0, unreadThreads: 0, isSubscribed: true,
+              myRights: {} } as unknown as Mailbox,
+          ];
+        },
+      }),
+    );
+    const mailboxes = await client.getAllMailboxes();
+    expect(mailboxes.map((m) => m.id)).toEqual(['mb1']);
+  });
+
+  it('keeps a REAL single-mailbox server result even after a transport failure', async () => {
+    // A genuine server that happens to return one inbox has a real id and real
+    // counts; only the exact placeholder shape may be replaced.
+    const real = {
+      id: 'real-inbox-id', name: 'Inbox', role: 'inbox', sortOrder: 0, totalEmails: 12,
+      unreadEmails: 1, totalThreads: 12, unreadThreads: 1, isSubscribed: true, myRights: {},
+    } as unknown as Mailbox;
+    const client = withOfflineFallback(
+      stubClient({ getAllMailboxes: async () => { noteTransportFailure(); return [real]; } }),
+    );
+    expect((await client.getAllMailboxes())[0].id).toBe('real-inbox-id');
+    expect(readOfflineMailboxes).not.toHaveBeenCalled();
+  });
+
+  it('never answers a read scoped to a delegated account', async () => {
+    // v1 replicates the PRIMARY mail account only, so the replica has no rows for
+    // a shared account and answering "empty" would be worse than the client's own.
+    const client = withOfflineFallback(
+      stubClient({
+        getEmails: async () => { noteTransportFailure(); return { emails: [], hasMore: false, total: 0 }; },
+      }),
+    );
+    await client.getEmails('inbox', 'someone-elses-account');
+    expect(readOfflineList).not.toHaveBeenCalled();
+  });
+
+  it('never answers a keyword- or category-filtered query', async () => {
+    // Those are server-side queries the replica does not reproduce. Serving an
+    // unfiltered page in their place would silently show the wrong set.
+    const failing = async () => { noteTransportFailure(); return { emails: [], hasMore: false, total: 0 }; };
+    const c1 = withOfflineFallback(stubClient({ getEmails: failing }));
+    await c1.getEmails('inbox', undefined, 25, 0, '$flagged');
+    expect(readOfflineList).not.toHaveBeenCalled();
+
+    resetTransportHealth();
+    const c2 = withOfflineFallback(stubClient({ getEmails: failing }));
+    await c2.getEmails('inbox', undefined, 25, 0, undefined, true, { from: 'x' });
+    expect(readOfflineList).not.toHaveBeenCalled();
+  });
+
+  it('stops asking once the replica reports itself absent', async () => {
+    isReplicaUnavailable.mockReturnValue(true);
+    const client = withOfflineFallback(
+      stubClient({ getEmail: async () => { noteTransportFailure(); return null; } }),
+    );
+    expect(await client.getEmail('e1')).toBeNull();
+    expect(readOfflineMessage).not.toHaveBeenCalled();
+  });
+
+  it('is idempotent, so re-wrapping a client does not stack fallbacks', async () => {
+    readOfflineMessage.mockResolvedValue({ email: replicaEmail('e1'), hasBody: true });
+    const base = stubClient({ getEmail: async () => { noteTransportFailure(); return null; } });
+    const once = withOfflineFallback(base);
+    const twice = withOfflineFallback(once);
+    expect(twice).toBe(once);
+    await twice.getEmail('e1');
+    expect(readOfflineMessage).toHaveBeenCalledTimes(1);
+  });
+});
diff --git a/lib/jmap/client.ts b/lib/jmap/client.ts
index 16502ccd..0f06a833 100644
--- a/lib/jmap/client.ts
+++ b/lib/jmap/client.ts
@@ -3,6 +3,7 @@ import type { SieveScript, SieveCapabilities } from "./sieve-types";
 import type { IJMAPClient } from "./client-interface";
 import { toWildcardQuery } from "./search-utils";
 import { batched, itemsPerRequest } from "./request-limits";
+import { noteTransportFailure, noteTransportSuccess } from "./transport-health";
 import { debug } from "@/lib/debug";
 import { normalizeCalendarEventLike } from "@/lib/calendar-event-normalization";
 
@@ -683,11 +684,24 @@ export class JMAPClient implements IJMAPClient {
     try {
       response = await fetch(url, { ...init, headers });
     } catch (error) {
+      // A `fetch` REJECTION - and only that - is a transport failure. Recorded so
+      // the offline replica's read fallback can tell "the network is down" from
+      // "the folder is empty", which the error-swallowing in getEmails/getEmail/
+      // getMailboxes otherwise makes indistinguishable (see
+      // lib/jmap/transport-health.ts). Deliberately NOT recorded for a 4xx/5xx or
+      // a 429: in those cases the server answered, so it is reachable.
+      noteTransportFailure();
       // Network error: retry once after brief delay (transient proxy/connection issues)
       if (this.reconnecting) throw error;
       await new Promise(r => setTimeout(r, 1000));
-      response = await fetch(url, { ...init, headers });
+      try {
+        response = await fetch(url, { ...init, headers });
+      } catch (retryError) {
+        noteTransportFailure();
+        throw retryError;
+      }
     }
+    noteTransportSuccess();
 
     // Handle 429 rate limiting - stop immediately, do not retry
     if (response.status === 429) {
diff --git a/lib/jmap/transport-health.ts b/lib/jmap/transport-health.ts
new file mode 100644
index 00000000..6b47ce9a
--- /dev/null
+++ b/lib/jmap/transport-health.ts
@@ -0,0 +1,62 @@
+// A single monotonic counter of JMAP TRANSPORT failures.
+//
+// WHY THIS EXISTS. The offline replica is a read-path FALLBACK, and to be one it
+// has to know that a read genuinely failed. `lib/jmap/client.ts` makes that
+// impossible to see from the outside: its read methods swallow their own errors
+// and return plausible-looking success. `getEmails()` returns
+// `{ emails: [], hasMore: false, total: 0 }`, so a dead network is
+// indistinguishable from an empty folder. `getEmail()` returns `null`.
+// `getMailboxes()` returns a SYNTHETIC single Inbox. Falling back on those shapes
+// alone would mean serving stale replica rows for a folder the user had genuinely
+// just emptied.
+//
+// So `authenticatedFetch` bumps this counter when, and only when, `fetch` itself
+// rejects - not on a 4xx, not on a 429 (that is a rate limit, and the server is
+// plainly reachable), not on a JMAP method error. The fallback layer samples the
+// counter before and after a call: a suspicious result PLUS an increment during
+// that exact call is a transport failure. Either signal alone is not enough.
+//
+// Module-level rather than per-client on purpose: it answers "is the network
+// working right now", which is a property of the machine, not of one account's
+// client instance.
+
+let failures = 0;
+let lastFailureAt = 0;
+let lastSuccessAt = 0;
+
+/** Called only when `fetch` itself rejects. Never for an HTTP status. */
+export function noteTransportFailure(): void {
+  failures++;
+  lastFailureAt = Date.now();
+}
+
+export function noteTransportSuccess(): void {
+  lastSuccessAt = Date.now();
+}
+
+/** Monotonic. Sample before and after a call to attribute a failure to it. */
+export function transportFailureCount(): number {
+  return failures;
+}
+
+export function transportHealth(): {
+  failures: number;
+  lastFailureAt: number;
+  lastSuccessAt: number;
+  /** Best-effort "probably offline": a failure more recent than any success. */
+  likelyOffline: boolean;
+} {
+  return {
+    failures,
+    lastFailureAt,
+    lastSuccessAt,
+    likelyOffline: lastFailureAt > lastSuccessAt,
+  };
+}
+
+/** Test-only reset. */
+export function resetTransportHealth(): void {
+  failures = 0;
+  lastFailureAt = 0;
+  lastSuccessAt = 0;
+}
diff --git a/lib/mail-index/store.ts b/lib/mail-index/store.ts
index 4e89793d..cae2a6f8 100644
--- a/lib/mail-index/store.ts
+++ b/lib/mail-index/store.ts
@@ -178,6 +178,12 @@ export class MailIndex {
     try {
       db.pragma('journal_mode = WAL');
       db.pragma('synchronous = NORMAL');
+      // The offline replica (lib/offline-replica/**) is a SECOND connection to
+      // this same file, writing disjoint tables. WAL lets a writer and readers
+      // coexist, but two WRITERS get SQLITE_BUSY immediately without this - and
+      // both subsystems are driven by the same renderer push handler, so they
+      // genuinely do overlap.
+      db.pragma('busy_timeout = 8000');
       version = readSchemaVersion(db);
     } catch {
       db.close();
@@ -189,6 +195,12 @@ export class MailIndex {
       assertEncrypted(db, dbPath);
       db.pragma('journal_mode = WAL');
       db.pragma('synchronous = NORMAL');
+      // The offline replica (lib/offline-replica/**) is a SECOND connection to
+      // this same file, writing disjoint tables. WAL lets a writer and readers
+      // coexist, but two WRITERS get SQLITE_BUSY immediately without this - and
+      // both subsystems are driven by the same renderer push handler, so they
+      // genuinely do overlap.
+      db.pragma('busy_timeout = 8000');
       version = null;
     }
 
diff --git a/lib/offline-fallback-client.ts b/lib/offline-fallback-client.ts
new file mode 100644
index 00000000..095e25a5
--- /dev/null
+++ b/lib/offline-fallback-client.ts
@@ -0,0 +1,180 @@
+// The read-path fallback. Wraps an `IJMAPClient` so that when a mail read fails
+// because the network is down, the answer comes from the encrypted offline replica
+// instead of an empty list.
+//
+// WHY THIS SHAPE, AND NOT A CACHE. The replica is consulted ONLY after a read has
+// genuinely failed at the transport level. That ordering is the whole coherence
+// story for the design review's H3: the webmail does local delta arithmetic on
+// mailbox unread counts for mark-read/move/delete, and if the replica sat in FRONT
+// of the server that arithmetic would operate on replica numbers and need
+// reconciliation rules. Behind the server, an online session never sees a replica
+// value at all, and while offline any count drift is bounded and repaired by the
+// next `Mailbox/changes`.
+//
+// WHY IT IS NOT ENOUGH TO LOOK AT THE RESULT. `lib/jmap/client.ts`'s read methods
+// swallow their own errors and return plausible success: `getEmails()` returns an
+// empty page, `getEmail()` returns `null`, `getMailboxes()` returns a SYNTHETIC
+// single Inbox. Falling back on those shapes alone would serve stale replica rows
+// for a folder the user had genuinely just emptied. So the test is TWO-PART: a
+// suspicious result AND a `fetch` rejection recorded during that exact call
+// (`lib/jmap/transport-health.ts`). A 4xx, a 429 or a JMAP method error all mean
+// the server answered, so none of them triggers a fallback.
+//
+// Mutates the instance rather than wrapping it in a Proxy: `JMAPClient` is a large
+// class whose methods call each other through `this`, and instance patching keeps
+// `this` identity exactly as it was. Idempotent, so re-wrapping the same client is
+// harmless.
+
+import { generateAccountId } from '@/lib/account-utils';
+import type { IJMAPClient } from '@/lib/jmap/client-interface';
+import type { Email, Mailbox } from '@/lib/jmap/types';
+import { transportFailureCount } from '@/lib/jmap/transport-health';
+import {
+  isReplicaUnavailable, readOfflineList, readOfflineMailboxes, readOfflineMessage,
+} from '@/lib/offline-replica-client';
+
+const WRAPPED = Symbol.for('vncmail.offlineFallback.wrapped');
+
+/**
+ * A `getMailboxes()` result that is really the client's offline placeholder.
+ *
+ * `client.ts` fabricates exactly this on failure: one mailbox, id `INBOX`, role
+ * `inbox`, zero counts. Matching it precisely matters - a real server that happens
+ * to return a single inbox has a real id and real counts.
+ */
+function isSyntheticMailboxList(mailboxes: readonly Mailbox[]): boolean {
+  return (
+    mailboxes.length === 1 &&
+    mailboxes[0]?.id === 'INBOX' &&
+    mailboxes[0]?.totalEmails === 0 &&
+    mailboxes[0]?.unreadEmails === 0
+  );
+}
+
+/** Resolves this client's cookie slot, so a multi-account shell reads the right replica. */
+async function slotFor(client: IJMAPClient): Promise<number | undefined> {
+  try {
+    const { useAccountStore } = await import('@/stores/account-store');
+    const id = generateAccountId(client.getUsername(), client.getServerUrl());
+    const accounts = useAccountStore.getState().accounts;
+    const match =
+      accounts.find((a) => a.id === id) ??
+      accounts.find((a) => a.serverIdentifiers?.includes(id));
+    return match?.cookieSlot;
+  } catch {
+    return undefined;
+  }
+}
+
+/**
+ * True when the replica may answer for this call.
+ *
+ * v1 replicates the PRIMARY mail account only, so a read explicitly scoped to a
+ * delegated/shared account must never be answered from it - the replica simply has
+ * no rows, and answering "empty" would be worse than the client's own empty.
+ */
+function scopedToPrimary(client: IJMAPClient, accountId?: string): boolean {
+  if (!accountId) return true;
+  try {
+    return accountId === client.getAccountId();
+  } catch {
+    return false;
+  }
+}
+
+export function withOfflineFallback<T extends IJMAPClient>(client: T): T {
+  const flagged = client as unknown as Record<symbol, boolean | undefined>;
+  if (flagged[WRAPPED]) return client;
+  flagged[WRAPPED] = true;
+
+  const target = client as unknown as IJMAPClient;
+  const originalGetEmail = target.getEmail.bind(target);
+  const originalGetEmails = target.getEmails.bind(target);
+  const originalGetMailboxes = target.getMailboxes.bind(target);
+  const originalGetAllMailboxes = target.getAllMailboxes.bind(target);
+
+  target.getEmail = async (emailId: string, accountId?: string): Promise<Email | null> => {
+    const before = transportFailureCount();
+    const online = await originalGetEmail(emailId, accountId);
+    if (online) return online;
+    if (isReplicaUnavailable()) return online;
+    // `null` alone is ambiguous: it is also what a genuinely-missing id returns.
+    // Only a transport failure during THIS call earns a fallback.
+    if (transportFailureCount() === before) return online;
+    if (!scopedToPrimary(client, accountId)) return online;
+
+    const offline = await readOfflineMessage(emailId, await slotFor(client));
+    // An envelope with no body would render blank AND leave the viewer's
+    // `isBodyLoading` gate stuck, so it is not an answer - better to keep the
+    // client's `null` and let the UI say the message is unavailable offline.
+    if (!offline?.email || !offline.hasBody) return online;
+    return offline.email;
+  };
+
+  target.getEmails = async (
+    mailboxId?: string,
+    accountId?: string,
+    limit: number = 50,
+    position: number = 0,
+    hasKeyword?: string,
+    pinnedFirst?: boolean,
+    extraFilter?: Record<string, unknown>,
+  ): Promise<{ emails: Email[]; hasMore: boolean; total: number }> => {
+    const before = transportFailureCount();
+    const online = await originalGetEmails(
+      mailboxId, accountId, limit, position, hasKeyword, pinnedFirst, extraFilter,
+    );
+    if (online.emails.length > 0) return online;
+    if (isReplicaUnavailable()) return online;
+    if (transportFailureCount() === before) return online;
+    if (!scopedToPrimary(client, accountId)) return online;
+    // A keyword or category filter is a server-side query the replica does not
+    // reproduce. Serving an unfiltered page in its place would silently show the
+    // wrong set, which is worse than showing nothing.
+    if (hasKeyword || extraFilter) return online;
+
+    const offline = await readOfflineList(mailboxId ?? null, {
+      limit,
+      offset: position,
+      slot: await slotFor(client),
+    });
+    if (!offline || offline.emails.length === 0) return online;
+    return { emails: offline.emails, hasMore: offline.hasMore, total: offline.total };
+  };
+
+  const mailboxFallback = async (
+    online: Mailbox[],
+    before: number,
+    accountId?: string,
+  ): Promise<Mailbox[]> => {
+    // Bail out unless the result is EMPTY or is the exact synthetic placeholder.
+    // Testing `length > 1` here was a real bug found by
+    // `lib/__tests__/offline-fallback-client.test.ts`: a server that legitimately
+    // exposes a single mailbox got its real folder - real id, real counts -
+    // replaced by replica rows the moment any unrelated transport blip was
+    // recorded during the call.
+    if (online.length > 0 && !isSyntheticMailboxList(online)) return online;
+    if (isReplicaUnavailable()) return online;
+    if (transportFailureCount() === before) return online;
+    if (!scopedToPrimary(client, accountId)) return online;
+    const offline = await readOfflineMailboxes(await slotFor(client));
+    if (!offline || offline.length === 0) return online;
+    return offline;
+  };
+
+  target.getMailboxes = async (accountId?: string): Promise<Mailbox[]> => {
+    const before = transportFailureCount();
+    const online = await originalGetMailboxes(accountId);
+    return mailboxFallback(online, before, accountId);
+  };
+
+  target.getAllMailboxes = async (): Promise<Mailbox[]> => {
+    const before = transportFailureCount();
+    const online = await originalGetAllMailboxes();
+    // `getAllMailboxes` falls back internally to `getMailboxes()`, so an offline
+    // run arrives here as the synthetic single Inbox rather than an empty list.
+    return mailboxFallback(online, before);
+  };
+
+  return client;
+}
diff --git a/lib/offline-replica-client.ts b/lib/offline-replica-client.ts
new file mode 100644
index 00000000..ba32c8b8
--- /dev/null
+++ b/lib/offline-replica-client.ts
@@ -0,0 +1,271 @@
+// Renderer-side client for the offline mail replica.
+//
+// The replica is EVENT-DRIVEN, exactly like the search index next to it: the
+// renderer already holds the live JMAP push connection, so a `StateChange` is what
+// triggers a sync cycle. There is no polling loop and no background worker.
+//
+// One cycle is BOUNDED (see lib/offline-replica/sync.ts's BUDGET), so a first
+// sync of a large mailbox needs several. `unfinishedWork` is the server saying
+// "call again", and `chainSync` below does that with a hard cap - the cap matters,
+// because an "unfinished work" signal that is true for a condition the cycle
+// cannot change is how the mobile client ended up chaining a new cycle every five
+// seconds forever.
+//
+// Every function here is best-effort and never throws: offline storage failing to
+// update must never break the mail UI.
+
+import { apiFetch } from '@/lib/browser-navigation';
+import { debug } from '@/lib/debug';
+import type { Email, Mailbox } from '@/lib/jmap/types';
+import type { StateChange } from '@/lib/jmap/types';
+
+export interface RetentionPolicy {
+  envelopeDays: number;
+  bodyDays: number;
+  maxBodyMB: number;
+}
+
+export interface CycleReport {
+  ok: boolean;
+  unfinishedWork: boolean;
+  bootstrapped: boolean;
+  reconciled: boolean;
+  mailboxesWritten: number;
+  envelopesWritten: number;
+  envelopesDeleted: number;
+  bodiesWritten: number;
+  bodiesEvicted: number;
+  coveragePhase: string;
+  resyncRequired: boolean;
+  warnings: string[];
+  errorClass?: string;
+  error?: string;
+  durationMs: number;
+}
+
+export interface ReplicaStats {
+  mailboxes: number;
+  envelopes: number;
+  bodies: number;
+  bodyBytes: number;
+  wantedBodies: number;
+  giveUps: number;
+  newest: string | null;
+  oldest: string | null;
+  fileBytes: number;
+}
+
+export interface ReplicaStatus {
+  ok: boolean;
+  policy: RetentionPolicy;
+  limits: Record<keyof RetentionPolicy, { min: number; max: number }>;
+  synced: boolean;
+  stats: ReplicaStats | null;
+  coveragePhase: string;
+  coveredFrom?: string | null;
+  resyncRequired: boolean;
+  lastCycleAt: number | null;
+  lastCycleOk: boolean | null;
+  lastCycleError?: string | null;
+}
+
+/** Set once the server says the feature isn't there, so we stop asking. */
+let knownUnavailable = false;
+let inFlight: Promise<CycleReport | null> | null = null;
+
+function slotQuery(slot?: number, extra?: string): string {
+  const params = new URLSearchParams();
+  if (typeof slot === 'number') params.set('slot', String(slot));
+  const base = params.toString();
+  if (extra && base) return `?${base}&${extra}`;
+  if (extra) return `?${extra}`;
+  return base ? `?${base}` : '';
+}
+
+/** True when the replica is known to be absent (not the desktop shell, or gated off). */
+export function isReplicaUnavailable(): boolean {
+  return knownUnavailable;
+}
+
+export function resetReplicaAvailability(): void {
+  knownUnavailable = false;
+}
+
+/**
+ * Runs ONE cycle. Single-flighted on the renderer as well as the server, so a
+ * burst of deliveries coalesces instead of queueing N overlapping requests that
+ * the server would then serialise anyway.
+ */
+export async function syncOnce(
+  opts: { slot?: number; policy?: RetentionPolicy; forceResync?: boolean } = {},
+): Promise<CycleReport | null> {
+  if (knownUnavailable) return null;
+  if (inFlight) return inFlight;
+
+  const run = (async (): Promise<CycleReport | null> => {
+    try {
+      const response = await apiFetch(`/api/offline/sync${slotQuery(opts.slot)}`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ policy: opts.policy, forceResync: opts.forceResync === true }),
+      });
+      // 404 = not the desktop shell. Permanent for this page load; stop asking so
+      // a busy mailbox doesn't post per delivery.
+      if (response.status === 404) { knownUnavailable = true; return null; }
+      if (response.status === 503) {
+        const body = await response.json().catch(() => ({}));
+        // A transport-class 503 means the BACKEND is unreachable, which is normal
+        // and temporary - it must not latch the feature off for the session. Only
+        // a missing binding / key channel does that.
+        const code = typeof body?.code === 'string' ? body.code : '';
+        if (code === 'no-binding' || code === 'no-key-channel' || code === 'unavailable') {
+          knownUnavailable = true;
+          return null;
+        }
+        return null;
+      }
+      if (!response.ok) return null;
+      const body = await response.json();
+      debug.log('push', '[replica] cycle', body?.report);
+      return (body?.report ?? null) as CycleReport | null;
+    } catch {
+      return null;
+    } finally {
+      inFlight = null;
+    }
+  })();
+
+  inFlight = run;
+  return run;
+}
+
+/** Hard cap on chained cycles per trigger. */
+export const MAX_CHAINED_CYCLES = 12;
+
+/**
+ * Runs cycles while the server reports unfinished work.
+ *
+ * The cap is the whole point. `unfinishedWork` is a hint, and a hint that stays
+ * true for something the cycle cannot resolve turns into an endless chain - which
+ * is exactly what happened on the mobile client when a body-queue counter reported
+ * attempted rather than inserted rows. The server-side fixes make that
+ * self-terminating; this cap means even a future regression costs a bounded number
+ * of requests rather than an infinite loop.
+ */
+export async function chainSync(
+  opts: { slot?: number; max?: number; onReport?: (report: CycleReport) => void } = {},
+): Promise<CycleReport | null> {
+  const max = Math.min(opts.max ?? MAX_CHAINED_CYCLES, MAX_CHAINED_CYCLES);
+  let last: CycleReport | null = null;
+  for (let i = 0; i < max; i++) {
+    const report = await syncOnce({ slot: opts.slot });
+    if (!report) return last;
+    last = report;
+    opts.onReport?.(report);
+    if (!report.ok || !report.unfinishedWork) return report;
+  }
+  return last;
+}
+
+/** The push-driven entry point. Fire-and-forget: the mail UI must not wait on it. */
+export function syncOnStateChange(change: StateChange, opts: { slot?: number } = {}): void {
+  if (knownUnavailable) return;
+  // Only mail-shaped changes are worth a cycle. A `Mailbox` state change alone is
+  // usually just an unread-count move, but the replica DOES hold those counts, so
+  // unlike the search index it is worth reacting to.
+  const relevant = Object.values(change.changed ?? {}).some(
+    (perAccount) => perAccount && (perAccount.Email || perAccount.Mailbox),
+  );
+  if (!relevant) return;
+  void syncOnce({ slot: opts.slot });
+}
+
+export async function fetchReplicaStatus(slot?: number): Promise<ReplicaStatus | null> {
+  try {
+    const response = await apiFetch(`/api/offline/status${slotQuery(slot)}`);
+    if (!response.ok) return null;
+    return (await response.json()) as ReplicaStatus;
+  } catch {
+    return null;
+  }
+}
+
+export async function updateRetentionPolicy(
+  policy: RetentionPolicy,
+  slot?: number,
+): Promise<boolean> {
+  try {
+    const response = await apiFetch(`/api/offline/status${slotQuery(slot)}`, {
+      method: 'PUT',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify(policy),
+    });
+    return response.ok;
+  } catch {
+    return false;
+  }
+}
+
+export async function purgeReplica(slot?: number): Promise<boolean> {
+  try {
+    const response = await apiFetch(`/api/offline/status${slotQuery(slot)}`, { method: 'DELETE' });
+    return response.ok;
+  } catch {
+    return false;
+  }
+}
+
+// ── reads ───────────────────────────────────────────────────────────────────
+
+interface ReadEnvelope<T> {
+  ok?: boolean;
+  available?: boolean;
+  error?: string;
+  data?: T;
+}
+
+async function read<T>(query: string): Promise<(T & { available: boolean }) | null> {
+  if (knownUnavailable) return null;
+  try {
+    const response = await apiFetch(`/api/offline/mail${query}`);
+    if (response.status === 404) { knownUnavailable = true; return null; }
+    if (!response.ok) return null;
+    const body = (await response.json()) as ReadEnvelope<unknown> & Record<string, unknown>;
+    if (body?.available !== true) return null;
+    return body as unknown as T & { available: boolean };
+  } catch {
+    return null;
+  }
+}
+
+export async function readOfflineMailboxes(slot?: number): Promise<Mailbox[] | null> {
+  const body = await read<{ mailboxes: Mailbox[] }>(slotQuery(slot, 'kind=mailboxes'));
+  return body?.mailboxes ?? null;
+}
+
+export async function readOfflineList(
+  mailboxId: string | null,
+  opts: { limit?: number; offset?: number; slot?: number } = {},
+): Promise<{ emails: Email[]; total: number; hasMore: boolean } | null> {
+  const params = new URLSearchParams({ kind: 'list' });
+  if (mailboxId !== null) params.set('mailboxId', mailboxId);
+  if (opts.limit !== undefined) params.set('limit', String(opts.limit));
+  if (opts.offset !== undefined) params.set('offset', String(opts.offset));
+  const body = await read<{ emails: Email[]; total: number; hasMore: boolean }>(
+    slotQuery(opts.slot, params.toString()),
+  );
+  if (!body) return null;
+  return { emails: body.emails ?? [], total: body.total ?? 0, hasMore: body.hasMore === true };
+}
+
+export async function readOfflineMessage(
+  id: string,
+  slot?: number,
+): Promise<{ email: Email | null; hasBody: boolean } | null> {
+  const params = new URLSearchParams({ kind: 'message', id });
+  const body = await read<{ email: Email | null; hasBody: boolean }>(
+    slotQuery(slot, params.toString()),
+  );
+  if (!body) return null;
+  return { email: body.email ?? null, hasBody: body.hasBody === true };
+}
diff --git a/lib/offline-replica/__tests__/apply.test.ts b/lib/offline-replica/__tests__/apply.test.ts
new file mode 100644
index 00000000..442a2776
--- /dev/null
+++ b/lib/offline-replica/__tests__/apply.test.ts
@@ -0,0 +1,126 @@
+import { describe, expect, it } from 'vitest';
+import {
+  advanceOneMs, madeForwardProgress, normalisePage, pageIsEmpty, planEmailFetches,
+  planMailboxFetches, updatedPropertiesAreCountsOnly, type ChangesPage,
+} from '../apply';
+import { asChangesState } from '../states';
+
+function page(partial: Partial<ChangesPage>): ChangesPage {
+  return {
+    oldState: asChangesState('old'),
+    newState: asChangesState('new'),
+    hasMoreChanges: false,
+    created: [],
+    updated: [],
+    destroyed: [],
+    ...partial,
+  };
+}
+
+describe('normalisePage', () => {
+  it('lets a destroyed id win outright over created and updated', () => {
+    // Fetching an id that is also destroyed spends a request to get `notFound`.
+    const out = normalisePage(page({ created: ['a', 'b'], updated: ['a'], destroyed: ['a'] }));
+    expect(out.created).toEqual(['b']);
+    expect(out.updated).toEqual([]);
+    expect(out.destroyed).toEqual(['a']);
+  });
+
+  it('treats an id in both created and updated as a create', () => {
+    // The create path fetches the full envelope tier, which already contains the
+    // updated values - so an extra 3-property fetch would be pure waste.
+    const out = normalisePage(page({ created: ['a'], updated: ['a'] }));
+    expect(out.created).toEqual(['a']);
+    expect(out.updated).toEqual([]);
+  });
+
+  it('deduplicates within each bucket', () => {
+    const out = normalisePage(page({ created: ['a', 'a'], destroyed: ['b', 'b'] }));
+    expect(out.created).toEqual(['a']);
+    expect(out.destroyed).toEqual(['b']);
+  });
+});
+
+describe('pageIsEmpty', () => {
+  it('is true only when nothing changed', () => {
+    // An empty page STILL has to advance the cursor: skipping it re-requests the
+    // same position forever.
+    expect(pageIsEmpty(page({}))).toBe(true);
+    expect(pageIsEmpty(page({ updated: ['a'] }))).toBe(false);
+  });
+});
+
+describe('planEmailFetches', () => {
+  it('drops an updated id we do not hold locally, BEFORE any fetch is issued', () => {
+    // The absent case is an unconditional no-op. Fetching it would need a
+    // `receivedAt` the 3-property response cannot supply and the schema's
+    // NOT NULL would reject. Coverage enumerates CURRENT state, so it will pick
+    // the record up with the updated values anyway.
+    const plan = planEmailFetches(page({ updated: ['have', 'missing'] }), new Set(['have']));
+    expect(plan.updateIds).toEqual(['have']);
+  });
+
+  it('keeps creates unconditional - presence is irrelevant for a create', () => {
+    const plan = planEmailFetches(page({ created: ['new'] }), new Set());
+    expect(plan.createIds).toEqual(['new']);
+  });
+
+  it('never routes an id into both the create and the update fetch', () => {
+    const plan = planEmailFetches(page({ created: ['a'], updated: ['a'] }), new Set(['a']));
+    expect(plan.createIds).toEqual(['a']);
+    expect(plan.updateIds).toEqual([]);
+  });
+});
+
+describe('updatedPropertiesAreCountsOnly', () => {
+  it('is true for the four counters and for an empty list', () => {
+    expect(updatedPropertiesAreCountsOnly(['unreadEmails'])).toBe(true);
+    expect(updatedPropertiesAreCountsOnly(['totalEmails', 'unreadThreads'])).toBe(true);
+    // "nothing but the state token moved" is counts-only vacuously.
+    expect(updatedPropertiesAreCountsOnly([])).toBe(true);
+  });
+
+  it('is false when the server will not say what changed', () => {
+    // `null` means "assume everything", so the whole object must be re-fetched.
+    expect(updatedPropertiesAreCountsOnly(null)).toBe(false);
+    expect(updatedPropertiesAreCountsOnly(undefined)).toBe(false);
+  });
+
+  it('is false as soon as one non-count property is present', () => {
+    expect(updatedPropertiesAreCountsOnly(['unreadEmails', 'name'])).toBe(false);
+  });
+});
+
+describe('planMailboxFetches', () => {
+  it('routes updates to the cheap four-integer patch when only counts moved', () => {
+    const plan = planMailboxFetches(
+      page({ created: ['new'], updated: ['old'], updatedProperties: ['unreadEmails'] }),
+    );
+    expect(plan.fullIds).toEqual(['new']);
+    expect(plan.countOnlyIds).toEqual(['old']);
+  });
+
+  it('re-fetches the whole object when updatedProperties is null', () => {
+    const plan = planMailboxFetches(page({ updated: ['old'], updatedProperties: null }));
+    expect(plan.fullIds).toEqual(['old']);
+    expect(plan.countOnlyIds).toEqual([]);
+  });
+});
+
+describe('keyset progress', () => {
+  it('requires STRICTLY greater, because `after` is spec-inclusive', () => {
+    // RFC 8621 s4.4.1: receivedAt "must be the same or after this date-time to
+    // match". So every page re-returns the boundary message, and equality is NOT
+    // progress - treating it as progress would loop on that millisecond forever.
+    expect(madeForwardProgress('2026-01-01T00:00:00.000Z', '2026-01-01T00:00:00.000Z')).toBe(false);
+    expect(madeForwardProgress('2026-01-01T00:00:00.001Z', '2026-01-01T00:00:00.000Z')).toBe(true);
+    expect(madeForwardProgress(null, '2026-01-01T00:00:00.000Z')).toBe(false);
+    expect(madeForwardProgress('2026-01-01T00:00:00.000Z', null)).toBe(true);
+  });
+
+  it('advances exactly one millisecond in the last-resort rung', () => {
+    expect(advanceOneMs('2026-01-01T00:00:00.000Z')).toBe('2026-01-01T00:00:00.001Z');
+    // A malformed value must not become NaN and poison the cursor.
+    expect(advanceOneMs('not-a-date')).toBe('not-a-date');
+  });
+});
diff --git a/lib/offline-replica/__tests__/errors.test.ts b/lib/offline-replica/__tests__/errors.test.ts
new file mode 100644
index 00000000..89033ccf
--- /dev/null
+++ b/lib/offline-replica/__tests__/errors.test.ts
@@ -0,0 +1,106 @@
+import { describe, expect, it } from 'vitest';
+import {
+  backoffDelayMs, classify, escalationApplies, movesCursor, nextRung, rungValue,
+  type ErrorClass,
+} from '../errors';
+
+const ALL: ErrorClass[] = [
+  'Transport', 'RateLimit', 'ServerTransient', 'RequestLimit', 'Auth', 'Fatal', 'StateInvalid',
+];
+
+describe('exactly one class moves a cursor', () => {
+  it('is StateInvalid, and nothing else', () => {
+    // This is the single load-bearing property of the taxonomy. Every other class
+    // leaves the cursor exactly where it was, which is what makes "a failure never
+    // causes silent data loss" structural rather than aspirational.
+    expect(ALL.filter(movesCursor)).toEqual(['StateInvalid']);
+  });
+
+  it('escalates to a rebuild only for size/availability problems', () => {
+    // Escalating on RateLimit would answer a rate-limited server with far MORE
+    // requests. On Auth, a 401 would trigger a rebuild. On Transport, a flaky
+    // tunnel would. Fatal is our own bug and a rebuild will not fix it.
+    expect(ALL.filter(escalationApplies).sort()).toEqual(['RequestLimit', 'ServerTransient']);
+  });
+});
+
+describe('classify', () => {
+  it('reads HTTP status before anything else', () => {
+    expect(classify({ httpStatus: 401 })).toBe('Auth');
+    expect(classify({ httpStatus: 403 })).toBe('Auth');
+    expect(classify({ httpStatus: 429 })).toBe('RateLimit');
+    expect(classify({ httpStatus: 413 })).toBe('RequestLimit');
+    expect(classify({ httpStatus: 503 })).toBe('ServerTransient');
+  });
+
+  it('classifies cannotCalculateChanges as the one cursor-moving class', () => {
+    expect(classify({ jmapErrorType: 'cannotCalculateChanges' })).toBe('StateInvalid');
+  });
+
+  it('defaults an UNRECOGNISED method error to ServerTransient', () => {
+    // Guessing transient costs a retry; guessing state-invalid costs a full
+    // resync; guessing fatal stalls the account. The cheapest wrong answer wins.
+    expect(classify({ jmapErrorType: 'somethingNobodyHasHeardOf' })).toBe('ServerTransient');
+  });
+
+  it('does not let a method error description masquerade as a transport failure', () => {
+    // Structure before strings: a method error's prose can legitimately contain
+    // "timeout" or "socket", and reading that as Transport would leave a genuine
+    // server-side problem being retried as though the network were down.
+    expect(classify({ jmapErrorType: 'invalidArguments', message: 'socket timeout' })).toBe('Fatal');
+  });
+
+  it('classifies a real fetch rejection as Transport', () => {
+    // "Offline is not an error": the cursor stands still and the work is retried.
+    for (const message of [
+      'fetch failed', 'connect ECONNREFUSED 127.0.0.1:1', 'getaddrinfo ENOTFOUND nope',
+      'socket hang up', 'The operation timed out',
+    ]) {
+      expect(classify({ message }), message).toBe('Transport');
+    }
+  });
+});
+
+describe('the maxChanges ladder is monotonically non-increasing for EVERY server value', () => {
+  it('never proposes a retry larger than the attempt that just failed', () => {
+    // Two historical bugs live here. An unbounded middle rung produced a retry
+    // STRICTLY LARGER than the failing attempt, actively worsening a
+    // "response too large" error. Clamping only rung 0 then reintroduced it in a
+    // narrower form: maxObjectsInGet=100 gave rung0=100 and rung1=250.
+    const serverValues = [
+      undefined, 1, 5, 10, 20, 25, 26, 49, 50, 51, 99, 100, 249, 250, 251, 499, 500, 501, 5000,
+    ];
+    for (const value of serverValues) {
+      const rungs = ([0, 1, 2, 3] as const).map((r) => rungValue(r, value));
+      for (let i = 1; i < rungs.length; i++) {
+        expect(
+          rungs[i],
+          `maxObjectsInGet=${value} rung ${i} (${rungs[i]}) must not exceed rung ${i - 1} (${rungs[i - 1]})`,
+        ).toBeLessThanOrEqual(rungs[i - 1]);
+      }
+      // And never zero, or the request asks for nothing and never progresses.
+      for (const r of rungs) expect(r).toBeGreaterThanOrEqual(1);
+    }
+  });
+
+  it('clamps rung 0 to what the server allows', () => {
+    expect(rungValue(0, 100)).toBe(100);
+    expect(rungValue(0, 5000)).toBe(500);
+    expect(rungValue(0, undefined)).toBe(500);
+  });
+
+  it('saturates rather than running off the end of the ladder', () => {
+    expect(nextRung(0)).toBe(1);
+    expect(nextRung(3)).toBe(3);
+  });
+});
+
+describe('backoff', () => {
+  it('is full-jitter and bounded by the cap', () => {
+    for (let attempt = 0; attempt < 12; attempt++) {
+      const delay = backoffDelayMs(attempt, { baseMs: 1000, capMs: 60_000 });
+      expect(delay).toBeGreaterThanOrEqual(0);
+      expect(delay).toBeLessThanOrEqual(60_000);
+    }
+  });
+});
diff --git a/lib/offline-replica/__tests__/retention.test.ts b/lib/offline-replica/__tests__/retention.test.ts
new file mode 100644
index 00000000..4f8377f5
--- /dev/null
+++ b/lib/offline-replica/__tests__/retention.test.ts
@@ -0,0 +1,154 @@
+// The clock-jump guard, and the wipe it caused on the mobile client.
+//
+// The bug being regressed here is not hypothetical: its reproduction on the mobile
+// side returned 0 envelopes from a 4-envelope store. The guard DETECTED the jump,
+// held the old floor for one cycle - and persisted the JUMPED floor. The next
+// chained cycle seconds later computed a floor within seconds of the persisted
+// one, so the guard passed, the movement was classified as a NARROW, and every
+// envelope below a floor a year in the future was evicted. Unrecoverable, because
+// `coveredFrom` then claims the range complete and `/changes` cannot re-deliver
+// pre-existing mail.
+
+import { describe, expect, it } from 'vitest';
+import {
+  adjustForWindow, CLOCK_JUMP_GUARD_MS, computeFloors, floorMovement,
+  guardFloorAgainstClockJump,
+} from '../retention';
+
+const DAY = 24 * 60 * 60 * 1000;
+const T0 = Date.parse('2026-08-05T12:00:00.000Z');
+
+function iso(t: number): string {
+  return new Date(t).toISOString();
+}
+
+describe('computeFloors', () => {
+  it('never lets the body window be wider than the envelope window', () => {
+    // A body with no envelope is an orphan by construction, and the whole point of
+    // two tiers is envelopes being a superset of bodies.
+    const floors = computeFloors({ envelopeDays: 30, bodyDays: 365, maxBodyMB: 100 }, T0);
+    expect(floors.bodyFrom).toBe(floors.envelopeFrom);
+  });
+
+  it('turns the MB cap into bytes', () => {
+    expect(computeFloors({ envelopeDays: 1, bodyDays: 1, maxBodyMB: 2 }, T0).maxBodyBytes)
+      .toBe(2 * 1024 * 1024);
+  });
+});
+
+describe('guardFloorAgainstClockJump', () => {
+  it('adopts the computed floor when there is no history to compare against', () => {
+    const g = guardFloorAgainstClockJump(iso(T0), undefined);
+    expect(g.suppressed).toBe(false);
+    expect(g.evictionAllowed).toBe(true);
+    expect(g.envelopeFrom).toBe(iso(T0));
+  });
+
+  it('adopts an ordinary drift - a DST shift must not trip it', () => {
+    const g = guardFloorAgainstClockJump(iso(T0 + 60 * 60 * 1000), iso(T0));
+    expect(g.suppressed).toBe(false);
+    expect(g.evictionAllowed).toBe(true);
+  });
+
+  it('suppresses a jump larger than the guard and refuses to authorise deletion', () => {
+    const jumped = iso(T0 + 365 * DAY);
+    const g = guardFloorAgainstClockJump(jumped, iso(T0));
+    expect(g.suppressed).toBe(true);
+    expect(g.envelopeFrom).toBe(iso(T0));
+    // Suppressing the FLOOR is not the same as suppressing the DELETIONS the
+    // floor authorises. Both the retention eviction and the reconcile sweep read
+    // this bit.
+    expect(g.evictionAllowed).toBe(false);
+    expect(g.warning).toBeTruthy();
+  });
+
+  it('THE H2 REGRESSION: persists the floor it USED, not the one it rejected', () => {
+    // This one assertion is the whole fix. Persisting the computed value here is
+    // what legitimised the anomaly on the very next cycle.
+    const jumped = iso(T0 + 365 * DAY);
+    const g = guardFloorAgainstClockJump(jumped, iso(T0));
+    expect(g.nextLastWindowFloor).toBe(iso(T0));
+    expect(g.nextLastWindowFloor).not.toBe(jumped);
+  });
+
+  it('THE H2 REGRESSION: stays suppressed across MANY chained cycles', () => {
+    // The original bug only showed on the SECOND cycle, so a single-cycle test
+    // passes against the broken code. Chaining is what reproduces it.
+    const stored = iso(T0);
+    let lastWindowFloor: string | undefined = stored;
+    for (let cycle = 0; cycle < 20; cycle++) {
+      // The clock is a year ahead and creeping forward a few seconds per cycle,
+      // exactly as a chained sync would observe it.
+      const computed = iso(T0 + 365 * DAY + cycle * 5_000);
+      const g = guardFloorAgainstClockJump(computed, lastWindowFloor);
+      expect(g.suppressed, `cycle ${cycle} must stay suppressed`).toBe(true);
+      expect(g.evictionAllowed, `cycle ${cycle} must not authorise deletion`).toBe(false);
+      expect(g.envelopeFrom, `cycle ${cycle} must keep the original floor`).toBe(stored);
+      lastWindowFloor = g.nextLastWindowFloor;
+    }
+    // And after 20 cycles the remembered floor is still the trustworthy one, so
+    // no later cycle can classify it as a narrow and evict everything.
+    expect(lastWindowFloor).toBe(stored);
+    expect(floorMovement(lastWindowFloor, stored)).toBe('unchanged');
+    expect(adjustForWindow(floorMovement(lastWindowFloor, stored), stored).evictBelow)
+      .toBeUndefined();
+  });
+
+  it('suppresses a BACKWARD jump too', () => {
+    const g = guardFloorAgainstClockJump(iso(T0 - 365 * DAY), iso(T0));
+    expect(g.suppressed).toBe(true);
+    expect(g.evictionAllowed).toBe(false);
+  });
+
+  it('treats an explicit retention change as INTENT and applies it, eviction included', () => {
+    // The computed floor moves for two independent reasons - the clock changing
+    // and the SETTING changing - and guarding a setting change is wrong. Without
+    // this discriminator a Settings edit sits unapplied until something unrelated
+    // moves the floor again.
+    const widened = iso(T0 - 365 * DAY);
+    const g = guardFloorAgainstClockJump(widened, iso(T0), { policyChanged: true });
+    expect(g.suppressed).toBe(false);
+    expect(g.evictionAllowed).toBe(true);
+    expect(g.envelopeFrom).toBe(widened);
+    expect(g.nextLastWindowFloor).toBe(widened);
+  });
+
+  it('a genuine user NARROW still evicts', () => {
+    // The guard must not become a reason nothing is ever deleted.
+    const narrowed = iso(T0 + 20 * 60 * 60 * 1000);
+    const g = guardFloorAgainstClockJump(narrowed, iso(T0));
+    expect(g.evictionAllowed).toBe(true);
+    expect(floorMovement(iso(T0), g.envelopeFrom)).toBe('narrowed');
+    expect(adjustForWindow('narrowed', g.envelopeFrom).evictBelow).toBe(narrowed);
+  });
+
+  it('tolerates an unparseable stored floor without wedging', () => {
+    const g = guardFloorAgainstClockJump(iso(T0), 'garbage');
+    // Date.parse('garbage') is NaN, so the delta is not finite: adopt rather than
+    // suppress forever on a corrupt value.
+    expect(g.suppressed).toBe(false);
+  });
+
+  it('uses a threshold above a day so a leap second or NTP nudge is invisible', () => {
+    expect(CLOCK_JUMP_GUARD_MS).toBeGreaterThan(DAY);
+  });
+});
+
+describe('floorMovement / adjustForWindow', () => {
+  it('a LATER floor keeps less mail and means evict', () => {
+    expect(floorMovement(iso(T0), iso(T0 + DAY))).toBe('narrowed');
+    expect(adjustForWindow('narrowed', iso(T0 + DAY))).toEqual({ evictBelow: iso(T0 + DAY) });
+  });
+
+  it('an EARLIER floor means re-scan, NOT a resync', () => {
+    // A widen moves the target back and re-enters coverage scanning. The cursors
+    // are untouched - a widen is not a reason to rebuild.
+    expect(floorMovement(iso(T0), iso(T0 - DAY))).toBe('widened');
+    expect(adjustForWindow('widened', iso(T0 - DAY))).toEqual({ rescanFrom: iso(T0 - DAY) });
+  });
+
+  it('does nothing without a previous floor', () => {
+    expect(floorMovement(undefined, iso(T0))).toBe('unchanged');
+    expect(adjustForWindow('unchanged', iso(T0))).toEqual({});
+  });
+});
diff --git a/lib/offline-replica/__tests__/states.test.ts b/lib/offline-replica/__tests__/states.test.ts
new file mode 100644
index 00000000..2d372345
--- /dev/null
+++ b/lib/offline-replica/__tests__/states.test.ts
@@ -0,0 +1,99 @@
+import fs from 'node:fs';
+import path from 'node:path';
+import { describe, expect, it } from 'vitest';
+import {
+  asChangesState, asSnapshotState, coveragePhaseForCommitment, mintEnumerationCommitment,
+} from '../states';
+
+describe('state token certification', () => {
+  it('rejects everything a parsed JSON body could hand over that is not a token', () => {
+    // The brand certifies PROVENANCE; this check certifies SHAPE. Without it a
+    // `null` or a number could be laundered into something the engine treats as a
+    // cursor forever.
+    for (const bad of [null, undefined, 0, 1, '', {}, [], true]) {
+      expect(() => asChangesState(bad)).toThrow(TypeError);
+      expect(() => asSnapshotState(bad)).toThrow(TypeError);
+    }
+  });
+
+  it('accepts a non-empty string', () => {
+    expect(asChangesState('s1')).toBe('s1');
+    expect(asSnapshotState('s1')).toBe('s1');
+  });
+});
+
+describe('EnumerationCommitment', () => {
+  it('is constructible - the symbol tag must be a real runtime Symbol', () => {
+    // `declare const tag: unique symbol` is type-level only and emits no runtime
+    // value, so using it as a computed key throws ReferenceError the first time
+    // the mint runs. That mistake is in the superseded design document; this test
+    // is what catches it.
+    const commitment = mintEnumerationCommitment({
+      jmapAccountId: 'a',
+      snapshot: asSnapshotState('snap'),
+      targetFrom: '2026-01-01T00:00:00.000Z',
+      sweepFloor: '2026-01-01T00:00:00.000Z',
+      kind: 'bootstrap',
+    });
+    expect(commitment.snapshot).toBe('snap');
+    expect(commitment.kind).toBe('bootstrap');
+  });
+
+  it('maps its kind onto the coverage phase', () => {
+    const base = {
+      jmapAccountId: 'a',
+      snapshot: asSnapshotState('snap'),
+      targetFrom: 'x',
+      sweepFloor: 'x',
+    } as const;
+    expect(coveragePhaseForCommitment(mintEnumerationCommitment({ ...base, kind: 'bootstrap' })))
+      .toBe('scanning');
+    expect(coveragePhaseForCommitment(mintEnumerationCommitment({ ...base, kind: 'reconcile' })))
+      .toBe('reconciling');
+  });
+
+  it('does not export its tag, so no object literal elsewhere can forge the type', () => {
+    const source = fs.readFileSync(path.join(__dirname, '..', 'states.ts'), 'utf8');
+    expect(source).toContain("const enumerationCommitmentTag = Symbol('EnumerationCommitment')");
+    expect(source).not.toMatch(/export\s+(const|let)\s+enumerationCommitmentTag/);
+    // And it must be a real Symbol() call, not the type-only declaration form.
+    expect(source).not.toMatch(/declare\s+const\s+enumerationCommitmentTag/);
+  });
+});
+
+describe('cursor provenance is greppable, not just documented', () => {
+  const replicaDir = path.join(__dirname, '..');
+
+  function sourceFiles(): string[] {
+    return fs
+      .readdirSync(replicaDir)
+      .filter((f) => f.endsWith('.ts'))
+      .map((f) => path.join(replicaDir, f));
+  }
+
+  it('mints branded states ONLY in jmap.ts (the response parser)', () => {
+    // This is the rule the whole brand exists to enforce. The mobile client's
+    // defect D4 was a snapshot state adopted as a /changes cursor after a
+    // transient 503; a cast anywhere outside the parser is how that comes back.
+    for (const file of sourceFiles()) {
+      const base = path.basename(file);
+      if (base === 'states.ts' || base === 'jmap.ts') continue;
+      const source = fs.readFileSync(file, 'utf8');
+      expect(source, `${base} must not mint a ChangesState`).not.toMatch(/asChangesState\s*\(/);
+      expect(source, `${base} must not mint a SnapshotState`).not.toMatch(/asSnapshotState\s*\(/);
+      expect(source, `${base} must not cast to a branded state`).not.toMatch(
+        /as\s+(ChangesState|SnapshotState)\b/,
+      );
+    }
+  });
+
+  it('mints an EnumerationCommitment ONLY where an enumeration is actually started', () => {
+    // A commitment is a promise to enumerate. Minting one anywhere that does not
+    // then enumerate makes the seed path's teeth meaningless.
+    const callers = sourceFiles().filter((file) => {
+      if (path.basename(file) === 'states.ts') return false;
+      return /mintEnumerationCommitment\s*\(/.test(fs.readFileSync(file, 'utf8'));
+    });
+    expect(callers.map((f) => path.basename(f))).toEqual(['sync.ts']);
+  });
+});
diff --git a/lib/offline-replica/__tests__/store.test.ts b/lib/offline-replica/__tests__/store.test.ts
new file mode 100644
index 00000000..0bf20bc6
--- /dev/null
+++ b/lib/offline-replica/__tests__/store.test.ts
@@ -0,0 +1,502 @@
+// Store-level invariants, against a REAL SQLCipher file.
+//
+// Skipped wholesale when the optional native binding is not installed (that is a
+// normal state on a platform with no prebuild - see lib/mail-index/binding.ts), so
+// this file must never be the only proof of anything.
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { randomBytes } from 'node:crypto';
+import { afterEach, beforeEach, describe, expect, it } from 'vitest';
+import { isSqlcipherAvailable } from '@/lib/mail-index/binding';
+import { indexDbPath } from '@/lib/mail-index/paths';
+import { clampPolicy, DEFAULT_POLICY, ReplicaStore } from '../store';
+import { reconcileStamp } from '../sync';
+import { asChangesState, asSnapshotState, mintEnumerationCommitment } from '../states';
+import type { EnvelopeRow } from '../types';
+
+const ACCOUNT = 'alice@example.org';
+const JMAP = 'jmap-account-1';
+
+function envelope(id: string, receivedAt: string, extra: Partial<EnvelopeRow> = {}): EnvelopeRow {
+  return {
+    jmapAccountId: JMAP,
+    id,
+    threadId: `t-${id}`,
+    receivedAt,
+    size: 1000,
+    subject: `subject ${id}`,
+    preview: `preview ${id}`,
+    fromJson: JSON.stringify([{ email: 'sender@example.org' }]),
+    toJson: null,
+    ccJson: null,
+    blobId: `blob-${id}`,
+    hasAttachment: false,
+    keywordsJson: '{}',
+    mailboxIds: ['inbox'],
+    ...extra,
+  };
+}
+
+describe.skipIf(!isSqlcipherAvailable())('ReplicaStore', () => {
+  let storeDir: string;
+  let key: Buffer;
+  let store: ReplicaStore;
+
+  beforeEach(() => {
+    storeDir = fs.mkdtempSync(path.join(os.tmpdir(), 'vncmail-replica-test-'));
+    key = randomBytes(32);
+    store = ReplicaStore.open({ storeDir, accountId: ACCOUNT, key });
+  });
+
+  afterEach(() => {
+    store.close();
+    fs.rmSync(storeDir, { recursive: true, force: true });
+  });
+
+  it('writes into the SAME file as the search index, and it is really encrypted', () => {
+    // One encryption boundary, one key, one purge. And `PRAGMA key` is a silent
+    // no-op on a non-SQLCipher binding, so the header check is the only thing that
+    // catches a store that "works" while sitting on disk in cleartext.
+    expect(store.dbPath).toBe(indexDbPath(storeDir, ACCOUNT));
+    store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+    store.close();
+    const header = fs.readFileSync(store.dbPath).subarray(0, 15).toString('latin1');
+    expect(header).not.toBe('SQLite format 3');
+    const raw = Buffer.concat(
+      ['', '-wal', '-shm']
+        .map((s) => `${store.dbPath}${s}`)
+        .filter((f) => fs.existsSync(f))
+        .map((f) => fs.readFileSync(f)),
+    );
+    expect(raw.includes('subject e1')).toBe(false);
+    // Re-open so afterEach's close() is harmless.
+    store = ReplicaStore.open({ storeDir, accountId: ACCOUNT, key });
+  });
+
+  describe('cursor provenance at the storage layer', () => {
+    it('refuses to create a cursor from nowhere', () => {
+      // A cursor is born from seedCursor and nowhere else. Creating one in
+      // advanceCursor would be a silent cursor-from-nowhere - exactly what the
+      // branded types exist to make impossible.
+      expect(() => store.advanceCursor({ jmapAccountId: JMAP, type: 'Email' }, asChangesState('s1')))
+        .toThrow(/seed it first/);
+    });
+
+    it('writes the cursor AND the coverage row it justifies in one transaction', () => {
+      store.transaction(() => {
+        store.seedCursor(
+          { jmapAccountId: JMAP, type: 'Email' },
+          mintEnumerationCommitment({
+            jmapAccountId: JMAP,
+            snapshot: asSnapshotState('snap-1'),
+            targetFrom: '2026-01-01T00:00:00.000Z',
+            sweepFloor: '2026-01-01T00:00:00.000Z',
+            kind: 'bootstrap',
+          }),
+          1000,
+        );
+      });
+      expect(store.getCursor({ jmapAccountId: JMAP, type: 'Email' })?.state).toBe('snap-1');
+      const coverage = store.getCoverage(JMAP);
+      expect(coverage?.phase).toBe('scanning');
+      expect(coverage?.sweepFloor).toBe('2026-01-01T00:00:00.000Z');
+    });
+
+    it('rolls back a seed whose commitment is for the wrong account', () => {
+      expect(() =>
+        store.transaction(() => {
+          store.seedCursor(
+            { jmapAccountId: JMAP, type: 'Email' },
+            mintEnumerationCommitment({
+              jmapAccountId: 'someone-else',
+              snapshot: asSnapshotState('snap'),
+              targetFrom: 'x', sweepFloor: 'x', kind: 'bootstrap',
+            }),
+            1000,
+          );
+        }),
+      ).toThrow(/different JMAP account/);
+      expect(store.getCursor({ jmapAccountId: JMAP, type: 'Email' })).toBeNull();
+    });
+
+    it('advances a seeded cursor and keeps counters field-level', () => {
+      seed(store);
+      store.transaction(() => {
+        store.advanceCursor({ jmapAccountId: JMAP, type: 'Email' }, asChangesState('s2'));
+        store.patchCursor({ jmapAccountId: JMAP, type: 'Email' }, { consecutiveFailures: 3 });
+      });
+      const cursor = store.getCursor({ jmapAccountId: JMAP, type: 'Email' });
+      expect(cursor?.state).toBe('s2');
+      expect(cursor?.consecutiveFailures).toBe(3);
+      // A patch must not be able to rewrite `state` - only advance/seed can.
+      store.transaction(() => {
+        store.patchCursor({ jmapAccountId: JMAP, type: 'Email' }, { drainPending: true });
+      });
+      expect(store.getCursor({ jmapAccountId: JMAP, type: 'Email' })?.state).toBe('s2');
+    });
+  });
+
+  describe('envelope tier', () => {
+    it('does NOT reset has_body on an idempotent replay', () => {
+      // Otherwise a replayed page looks like "body missing" to the backfill job and
+      // re-downloads every body in the page.
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+      store.transaction(() => { store.putBodyIfEnvelopeExists(JMAP, 'e1', '{"bodyValues":{}}'); });
+      expect(store.envelopesWithoutBody(JMAP, '2026-01-01T00:00:00.000Z', 10)).toHaveLength(0);
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 2); });
+      expect(
+        store.envelopesWithoutBody(JMAP, '2026-01-01T00:00:00.000Z', 10),
+        'a replayed envelope upsert must not clear has_body',
+      ).toHaveLength(0);
+    });
+
+    it('patches only the two mutable properties, and no-ops for an absent id', () => {
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+      const ok = store.transaction(() =>
+        store.patchEnvelopeMutable(JMAP, 'e1', { keywordsJson: '{"$seen":true}', mailboxIds: ['archive'] }),
+      );
+      expect(ok).toBe(true);
+      expect(store.mailboxIdsFor(JMAP, 'e1')).toEqual(['archive']);
+      // An update for an id we do not hold must leave no membership rows behind.
+      const missing = store.transaction(() =>
+        store.patchEnvelopeMutable(JMAP, 'nope', { keywordsJson: '{}', mailboxIds: ['inbox'] }),
+      );
+      expect(missing).toBe(false);
+      expect(store.mailboxIdsFor(JMAP, 'nope')).toEqual([]);
+    });
+
+    it('never writes a body whose envelope is gone', () => {
+      // A body fetched moments before its envelope was destroyed in the same cycle
+      // would otherwise land as an orphan.
+      const wrote = store.transaction(() => store.putBodyIfEnvelopeExists(JMAP, 'ghost', '{}'));
+      expect(wrote).toBe(false);
+      expect(store.getBody(JMAP, 'ghost')).toBeNull();
+    });
+
+    it('deleting an email takes its body, membership and queue row with it', () => {
+      store.transaction(() => {
+        store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1);
+        store.enqueueBodies([{ emailId: 'e1', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 }]);
+      });
+      store.transaction(() => { store.putBodyIfEnvelopeExists(JMAP, 'e1', '{"a":1}'); });
+      store.transaction(() => { store.deleteEmails(JMAP, ['e1']); });
+      expect(store.getBody(JMAP, 'e1')).toBeNull();
+      expect(store.mailboxIdsFor(JMAP, 'e1')).toEqual([]);
+      expect(store.countWantedBodies(JMAP, Date.now())).toBe(0);
+    });
+  });
+
+  describe('the reconcile sweep', () => {
+    it('refuses to run without a pinned stamp rather than deleting unverified records', () => {
+      expect(() => store.sweep(JMAP, '2026-01-01T00:00:00.000Z', undefined))
+        .toThrow(/refusing to delete unverified/);
+    });
+
+    it('keeps what the enumeration re-saw and deletes what it did not', () => {
+      // The whole "seen set as one integer" trick: re-upserting refreshes
+      // cached_at, and the sweep deletes anything still below the pin.
+      store.transaction(() => {
+        store.upsertEnvelopes([
+          envelope('kept', '2026-08-01T00:00:00.000Z'),
+          envelope('gone', '2026-08-02T00:00:00.000Z'),
+        ], 100);
+      });
+      const stamp = Math.max(500, store.maxEnvelopeCachedAt(JMAP) + 1);
+      store.transaction(() => {
+        store.upsertEnvelopes([envelope('kept', '2026-08-01T00:00:00.000Z')], stamp);
+      });
+      store.transaction(() => { store.sweep(JMAP, '2026-07-01T00:00:00.000Z', stamp); });
+      expect(store.getEnvelopeRaw(JMAP, 'kept')).not.toBeNull();
+      expect(store.getEnvelopeRaw(JMAP, 'gone')).toBeNull();
+    });
+
+    it('a stamp taken from a FROZEN clock would sweep nothing; the derived one works', () => {
+      // Both halves matter, and both fail silently. Exercising the real
+      // `reconcileStamp` rather than re-deriving it in the test is the point.
+      store.transaction(() => {
+        store.upsertEnvelopes([
+          envelope('kept', '2026-08-01T00:00:00.000Z'),
+          envelope('gone', '2026-08-02T00:00:00.000Z'),
+        ], 9_999);
+      });
+      const frozenNow = 1_000;
+
+      // The naive version: with the clock behind the data, nothing is below the
+      // stamp, so a re-verified store sweeps zero rows and stale records live on.
+      expect(store.sweep(JMAP, '2026-07-01T00:00:00.000Z', frozenNow)).toBe(0);
+
+      const stamp = reconcileStamp(frozenNow, store.maxEnvelopeCachedAt(JMAP));
+      expect(stamp).toBe(10_000);
+      store.transaction(() => {
+        store.upsertEnvelopes([envelope('kept', '2026-08-01T00:00:00.000Z')], stamp);
+      });
+      expect(store.transaction(() => store.sweep(JMAP, '2026-07-01T00:00:00.000Z', stamp))).toBe(1);
+      expect(store.getEnvelopeRaw(JMAP, 'kept')).not.toBeNull();
+      expect(store.getEnvelopeRaw(JMAP, 'gone')).toBeNull();
+    });
+
+    it('stamping an enumeration with `now` instead of the pin deletes what it just verified', () => {
+      // The other direction of the same bug: the pin EXCEEDS now, so a page that
+      // stamps with `now` lands below the pin and the sweep eats it.
+      store.transaction(() => {
+        store.upsertEnvelopes([envelope('verified', '2026-08-01T00:00:00.000Z')], 9_999);
+      });
+      const now = 1_000;
+      const stamp = reconcileStamp(now, store.maxEnvelopeCachedAt(JMAP));
+      // Re-verified against the server, but stamped with the WRONG value.
+      store.transaction(() => {
+        store.upsertEnvelopes([envelope('verified', '2026-08-01T00:00:00.000Z')], now);
+      });
+      store.transaction(() => { store.sweep(JMAP, '2026-07-01T00:00:00.000Z', stamp); });
+      expect(
+        store.getEnvelopeRaw(JMAP, 'verified'),
+        'this is the failure mode the pinned stamp exists to prevent',
+      ).toBeNull();
+    });
+  });
+
+  describe('the body queue - the durable-terminal-state fixes', () => {
+    it('enqueueBodies reports rows ACTUALLY INSERTED, not attempted', () => {
+      // Reporting the attempted count made the mobile engine believe there was
+      // unfinished work every cycle for as long as any envelope lacked a body,
+      // chaining a new cycle every few seconds indefinitely.
+      const entry = { emailId: 'e1', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 };
+      expect(store.transaction(() => store.enqueueBodies([entry]))).toBe(1);
+      expect(store.transaction(() => store.enqueueBodies([entry]))).toBe(0);
+    });
+
+    it('never resets attempts on a re-enqueue', () => {
+      const entry = { emailId: 'e1', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 };
+      store.transaction(() => { store.enqueueBodies([entry]); });
+      store.transaction(() => { store.bumpBodyAttempt(JMAP, 'e1', 0, 'boom'); });
+      store.transaction(() => { store.enqueueBodies([entry]); });
+      expect(store.takeBodyQueue(JMAP, 10, Date.now())[0]?.attempts).toBe(1);
+    });
+
+    it('THE H1 REGRESSION: a gave-up row is KEPT and is never revived by a re-enqueue', () => {
+      // Deleting the row on give-up was not enough: the backfill driver is
+      // "envelope with no body", which cannot tell "not fetched yet" from
+      // "deliberately not kept", so the next pass re-inserted a fresh attempts=0
+      // row and a permanently-failing body was retried five times per cycle forever.
+      store.transaction(() => {
+        store.markBodyGaveUp(JMAP, [
+          { emailId: 'e1', receivedAt: '2026-08-01T00:00:00.000Z', reason: 'attempts' },
+        ]);
+      });
+      expect(store.listBodyGiveUps(JMAP, 10)).toEqual(['e1']);
+      // Not WANTED any more, so the drain never picks it up again.
+      expect(store.takeBodyQueue(JMAP, 10, Date.now())).toHaveLength(0);
+      // And a re-enqueue cannot resurrect it.
+      const inserted = store.transaction(() =>
+        store.enqueueBodies([
+          { emailId: 'e1', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 },
+        ]),
+      );
+      expect(inserted).toBe(0);
+      expect(store.takeBodyQueue(JMAP, 10, Date.now())).toHaveLength(0);
+    });
+
+    it('THE H1c REGRESSION: a cap-shed body is markable even with NO existing queue row', () => {
+      // The download/discard loop: the cap sheds a body that was fetched and stored
+      // successfully, so there is no queue row left to UPDATE. If the mark is
+      // silently dropped, the envelope is still inside the body WINDOW, the backfill
+      // re-enqueues it, it downloads again, and the cap sheds it again - unbounded
+      // data use that never terminates.
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+      store.transaction(() => { store.putBodyIfEnvelopeExists(JMAP, 'e1', '{"bodyValues":{"1":{"value":"x"}}}'); });
+      expect(store.takeBodyQueue(JMAP, 10, Date.now())).toHaveLength(0); // no queue row exists
+
+      store.transaction(() => {
+        store.deleteBodies(JMAP, ['e1']);
+        store.markBodyGaveUp(JMAP, [
+          { emailId: 'e1', receivedAt: '2026-08-01T00:00:00.000Z', reason: 'shed-by-cap' },
+        ]);
+      });
+      expect(
+        store.listBodyGiveUps(JMAP, 10),
+        'the cap-shed mark must be an upsert, or the shed/re-download loop stays open',
+      ).toEqual(['e1']);
+      // The envelope is back to has_body=0 and still in the window, so without the
+      // mark the backfill WOULD pick it up. With the mark it is excluded.
+      expect(store.envelopesWithoutBody(JMAP, '2026-01-01T00:00:00.000Z', 10).map((e) => e.id))
+        .toEqual(['e1']);
+      expect(store.listBodyGiveUps(JMAP, 10)).toContain('e1');
+    });
+
+    it('clearing give-ups DELETES them, so they look like "never queued"', () => {
+      // A cleared give-up must come back with a clean attempt count, which an
+      // un-flag would not give.
+      store.transaction(() => {
+        store.markBodyGaveUp(JMAP, [
+          { emailId: 'a', receivedAt: '2026-08-01T00:00:00.000Z', reason: 'attempts' },
+          { emailId: 'b', receivedAt: '2026-08-01T00:00:00.000Z', reason: 'shed-by-cap' },
+        ]);
+      });
+      store.transaction(() => { store.clearBodyGiveUps(JMAP, 'shed-by-cap'); });
+      expect(store.listBodyGiveUps(JMAP, 10)).toEqual(['a']);
+      store.transaction(() => { store.clearBodyGiveUps(JMAP); });
+      expect(store.listBodyGiveUps(JMAP, 10)).toEqual([]);
+      expect(
+        store.transaction(() =>
+          store.enqueueBodies([
+            { emailId: 'a', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 },
+          ]),
+        ),
+        'a cleared give-up must be re-enqueueable',
+      ).toBe(1);
+    });
+
+    it('honours a backoff window', () => {
+      store.transaction(() => {
+        store.enqueueBodies([
+          { emailId: 'e1', jmapAccountId: JMAP, receivedAt: '2026-08-01T00:00:00.000Z', attempts: 0 },
+        ]);
+      });
+      store.transaction(() => { store.bumpBodyAttempt(JMAP, 'e1', 10_000, 'later'); });
+      expect(store.takeBodyQueue(JMAP, 10, 5_000)).toHaveLength(0);
+      expect(store.takeBodyQueue(JMAP, 10, 20_000)).toHaveLength(1);
+    });
+  });
+
+  describe('eviction', () => {
+    it('cap eviction takes the oldest bodies first and leaves envelopes alone', () => {
+      store.transaction(() => {
+        store.upsertEnvelopes([
+          envelope('old', '2026-01-01T00:00:00.000Z'),
+          envelope('new', '2026-08-01T00:00:00.000Z'),
+        ], 1);
+      });
+      store.transaction(() => {
+        store.putBodyIfEnvelopeExists(JMAP, 'old', '{"v":"old"}');
+        store.putBodyIfEnvelopeExists(JMAP, 'new', '{"v":"new"}');
+      });
+      expect(store.oldestBodies(JMAP, 1).map((b) => b.emailId)).toEqual(['old']);
+      store.transaction(() => { store.deleteBodies(JMAP, ['old']); });
+      // The message stays LISTED - only its content went.
+      expect(store.getEnvelopeRaw(JMAP, 'old')).not.toBeNull();
+      expect(store.countBodies(JMAP)).toBe(1);
+    });
+
+    it('no deletion path leaves an orphan body behind', () => {
+      // This is the real invariant. `orphanBodies()` is a belt-and-braces sweep for
+      // orphans a CRASH between two transactions could leave; it is deliberately
+      // not reachable through the store's own API, which is what this asserts.
+      // (So the detection query itself is covered only by the integration run, not
+      // by this file - stated rather than papered over with a vacuous assertion.)
+      store.transaction(() => {
+        store.upsertEnvelopes([
+          envelope('a', '2026-01-01T00:00:00.000Z'),
+          envelope('b', '2026-08-01T00:00:00.000Z'),
+        ], 1);
+      });
+      store.transaction(() => {
+        store.putBodyIfEnvelopeExists(JMAP, 'a', '{"v":1}');
+        store.putBodyIfEnvelopeExists(JMAP, 'b', '{"v":2}');
+      });
+      expect(store.countBodies(JMAP)).toBe(2);
+
+      store.transaction(() => { store.deleteEmails(JMAP, ['a']); });
+      expect(store.orphanBodies(JMAP, 10)).toEqual([]);
+
+      store.transaction(() => { store.evictEnvelopesBelow(JMAP, '2026-09-01T00:00:00.000Z'); });
+      expect(store.countEnvelopes(JMAP)).toBe(0);
+      expect(store.countBodies(JMAP)).toBe(0);
+      expect(store.orphanBodies(JMAP, 10)).toEqual([]);
+    });
+  });
+
+  describe('purge', () => {
+    it('purgeAll takes the CURSORS with the records', () => {
+      // A record wipe that leaves a live cursor behind is the one state no amount
+      // of syncing repairs: /changes cannot re-deliver mail that already existed
+      // when the cursor was captured.
+      seed(store);
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+      store.transaction(() => { store.purgeAll(); });
+      expect(store.countEnvelopes(JMAP)).toBe(0);
+      expect(store.getCursor({ jmapAccountId: JMAP, type: 'Email' })).toBeNull();
+      expect(store.getCoverage(JMAP)).toBeNull();
+    });
+
+    it('a wrong key is treated as unreadable and rebuilt, never as a prompt', () => {
+      store.transaction(() => { store.upsertEnvelopes([envelope('e1', '2026-08-01T00:00:00.000Z')], 1); });
+      store.close();
+      const other = ReplicaStore.open({ storeDir, accountId: ACCOUNT, key: randomBytes(32) });
+      try {
+        expect(other.countEnvelopes(JMAP)).toBe(0);
+      } finally {
+        other.close();
+      }
+      store = ReplicaStore.open({ storeDir, accountId: ACCOUNT, key });
+    });
+  });
+
+  describe('policy', () => {
+    it('round-trips and clamps', () => {
+      store.transaction(() => { store.setPolicy({ envelopeDays: 99999, bodyDays: 0, maxBodyMB: 1 }); });
+      const policy = store.getPolicy();
+      expect(policy.envelopeDays).toBe(3650);
+      expect(policy.bodyDays).toBe(1);
+      expect(policy.maxBodyMB).toBe(16);
+    });
+
+    it('defaults when nothing was ever written', () => {
+      expect(store.getPolicy()).toEqual(DEFAULT_POLICY);
+    });
+  });
+
+  describe('read path', () => {
+    it('lists a mailbox page newest-first with a correct total', () => {
+      store.transaction(() => {
+        store.upsertEnvelopes([
+          envelope('a', '2026-08-01T00:00:00.000Z'),
+          envelope('b', '2026-08-02T00:00:00.000Z'),
+          envelope('c', '2026-08-03T00:00:00.000Z', { mailboxIds: ['archive'] }),
+        ], 1);
+      });
+      const inbox = store.listEnvelopes(JMAP, 'inbox', 10, 0);
+      expect(inbox.total).toBe(2);
+      expect(inbox.rows.map((r) => String(r.id))).toEqual(['b', 'a']);
+      // A null mailbox is "everything", which is what the unified views want.
+      expect(store.listEnvelopes(JMAP, null, 10, 0).total).toBe(3);
+      expect(store.listEnvelopes(JMAP, 'archive', 10, 0).rows.map((r) => String(r.id))).toEqual(['c']);
+    });
+
+    it('reports the account ids it holds without needing a network session', () => {
+      store.transaction(() => { store.upsertEnvelopes([envelope('a', '2026-08-01T00:00:00.000Z')], 1); });
+      expect(store.knownJmapAccountIds()).toEqual([JMAP]);
+    });
+  });
+});
+
+describe('clampPolicy', () => {
+  it('never lets the body window exceed the envelope window', () => {
+    expect(clampPolicy({ envelopeDays: 30, bodyDays: 365, maxBodyMB: 100 }).bodyDays).toBe(30);
+  });
+
+  it('falls back to defaults for junk input', () => {
+    expect(clampPolicy({ envelopeDays: NaN } as never).envelopeDays).toBe(DEFAULT_POLICY.envelopeDays);
+    expect(clampPolicy(null)).toEqual(DEFAULT_POLICY);
+    expect(clampPolicy(undefined)).toEqual(DEFAULT_POLICY);
+  });
+});
+
+function seed(store: ReplicaStore): void {
+  store.transaction(() => {
+    for (const type of ['Email', 'Mailbox'] as const) {
+      store.seedCursor(
+        { jmapAccountId: JMAP, type },
+        mintEnumerationCommitment({
+          jmapAccountId: JMAP,
+          snapshot: asSnapshotState('snap'),
+          targetFrom: '2026-01-01T00:00:00.000Z',
+          sweepFloor: '2026-01-01T00:00:00.000Z',
+          kind: 'bootstrap',
+        }),
+        1000,
+      );
+    }
+  });
+}
diff --git a/lib/offline-replica/apply.ts b/lib/offline-replica/apply.ts
new file mode 100644
index 00000000..12ee7225
--- /dev/null
+++ b/lib/offline-replica/apply.ts
@@ -0,0 +1,155 @@
+// Change-application planning. PURE: no network, no storage, no store access.
+//
+// That purity is the single highest-leverage constraint in the design, because
+// `plan(page, presentIds) -> what to fetch` is assertable as plain data. It is
+// what turns a failure-mode table into a test suite rather than a promise.
+
+import type { ChangesState } from './states';
+
+export interface ChangesPage {
+  oldState: ChangesState;
+  newState: ChangesState;
+  hasMoreChanges: boolean;
+  created: string[];
+  updated: string[];
+  destroyed: string[];
+  /** `Mailbox/changes` only. `null`/absent means "assume everything changed". */
+  updatedProperties?: string[] | null;
+}
+
+/**
+ * Collapses the overlap RFC 8620 permits, BEFORE anything iterates - so
+ * downstream code cannot get the order wrong by following the server's array
+ * order.
+ */
+export function normalisePage(page: ChangesPage): {
+  created: string[];
+  updated: string[];
+  destroyed: string[];
+} {
+  const destroyed = [...new Set(page.destroyed)];
+  const destroyedSet = new Set(destroyed);
+  // An id in `destroyed` wins outright: fetching it would be wasted and the
+  // result would be `notFound`.
+  const created = [...new Set(page.created)].filter((id) => !destroyedSet.has(id));
+  const createdSet = new Set(created);
+  // An id in both `created` and `updated` is a CREATE - the create path fetches
+  // the full envelope tier, which already includes the updated values.
+  const updated = [...new Set(page.updated)].filter(
+    (id) => !destroyedSet.has(id) && !createdSet.has(id),
+  );
+  return { created, updated, destroyed };
+}
+
+/**
+ * An empty page STILL ADVANCES THE CURSOR. Skipping it re-requests the same
+ * position forever.
+ */
+export function pageIsEmpty(page: ChangesPage): boolean {
+  return page.created.length === 0 && page.updated.length === 0 && page.destroyed.length === 0;
+}
+
+export interface EmailFetchPlan {
+  /** Full envelope tier. */
+  createIds: string[];
+  /** THREE properties only: id, keywords, mailboxIds. Never bodies. */
+  updateIds: string[];
+  destroyIds: string[];
+}
+
+/**
+ * `keywords` and `mailboxIds` are the ONLY mutable Email properties
+ * (RFC 8621 s4.1). Body structure, body values, attachments, headers,
+ * `receivedAt`, `size`, `threadId`, `preview`, `subject`, addresses and
+ * `hasAttachment` are all immutable for the lifetime of the id.
+ *
+ * So an `updated` Email cannot have a changed body, and re-fetching one is pure
+ * waste. This is also what stops a message cached while unread from staying
+ * unread forever.
+ *
+ * An `updated` id we do NOT hold locally is an UNCONDITIONAL NO-OP, filtered out
+ * BEFORE the fetch is issued. Cheaper, and it avoids having to fabricate a
+ * `receivedAt` that a 3-property response cannot supply and the schema's NOT NULL
+ * would reject. Safe to ignore because absence is always either "retention
+ * decided against it" or "coverage has not reached it yet" - and coverage
+ * enumerates CURRENT state, so it will pick the record up with the updated values
+ * anyway. Nothing needs the update replayed.
+ */
+export function planEmailFetches(
+  page: ChangesPage,
+  presentIds: ReadonlySet<string>,
+): EmailFetchPlan {
+  const { created, updated, destroyed } = normalisePage(page);
+  return {
+    createIds: created,
+    updateIds: updated.filter((id) => presentIds.has(id)),
+    destroyIds: destroyed,
+  };
+}
+
+const COUNT_PROPERTIES = new Set([
+  'totalEmails', 'unreadEmails', 'totalThreads', 'unreadThreads',
+]);
+
+/**
+ * True when a `Mailbox/changes` update touched only the four counters, so a
+ * four-integer patch is enough instead of re-fetching every folder object.
+ *
+ * `updatedProperties: null` means the server will not say, so everything must be
+ * re-fetched. An EMPTY array means "nothing but the state token moved", which is
+ * counts-only vacuously.
+ */
+export function updatedPropertiesAreCountsOnly(
+  updatedProperties: readonly string[] | null | undefined,
+): boolean {
+  if (!updatedProperties) return false;
+  if (updatedProperties.length === 0) return true;
+  return updatedProperties.every((p) => COUNT_PROPERTIES.has(p));
+}
+
+export interface MailboxFetchPlan {
+  /** Needs the whole object. */
+  fullIds: string[];
+  /** Only the count columns move. */
+  countOnlyIds: string[];
+  destroyIds: string[];
+}
+
+export function planMailboxFetches(page: ChangesPage): MailboxFetchPlan {
+  const { created, updated, destroyed } = normalisePage(page);
+  const countsOnly = updatedPropertiesAreCountsOnly(page.updatedProperties);
+  return {
+    fullIds: countsOnly ? created : [...created, ...updated],
+    countOnlyIds: countsOnly ? updated : [],
+    destroyIds: destroyed,
+  };
+}
+
+/**
+ * Keyset-walk progress test.
+ *
+ * `after` is INCLUSIVE - this is specified, not implementation-defined.
+ * RFC 8621 s4.4.1: the `receivedAt` of the Email "must be the same or after this
+ * date-time to match the condition". So every page after the first re-returns the
+ * boundary message(s); dedupe by id on commit makes that free. But forward
+ * progress therefore requires `max(receivedAt)` STRICTLY GREATER than the cursor.
+ *
+ * Treating `after` as exclusive and adding a millisecond, as an earlier revision
+ * of the mobile design did, silently skips every message sharing the boundary
+ * millisecond on any conforming server.
+ */
+export function madeForwardProgress(
+  maxReceivedAt: string | null,
+  scanCursor: string | null,
+): boolean {
+  if (maxReceivedAt === null) return false;
+  if (scanCursor === null) return true;
+  return maxReceivedAt > scanCursor;
+}
+
+/** Advance a scan cursor by exactly one millisecond. The last-resort paging rung. */
+export function advanceOneMs(iso: string): string {
+  const t = Date.parse(iso);
+  if (!Number.isFinite(t)) return iso;
+  return new Date(t + 1).toISOString();
+}
diff --git a/lib/offline-replica/engine.ts b/lib/offline-replica/engine.ts
new file mode 100644
index 00000000..e00a94cb
--- /dev/null
+++ b/lib/offline-replica/engine.ts
@@ -0,0 +1,190 @@
+// Session resolution, store lifecycle and single-flight. The thin layer every
+// `/api/offline/*` replica route goes through.
+
+import { logger } from '@/lib/logger';
+import { fetchJmapSession, accountIdFor, CAP_MAIL, type JmapSessionInfo } from '@/lib/mail-index/jmap';
+import { withIndexKey } from '@/lib/mail-index/key';
+import { getStoreDir } from '@/lib/mail-index/paths';
+import {
+  IndexSessionError, resolveIndexSession, type IndexSession,
+} from '@/lib/mail-index/reindex';
+import { classify, ReplicaSyncError } from './errors';
+import { ReplicaStore, type RetentionPolicy } from './store';
+import { BUDGET, runCycle, type CycleReport } from './sync';
+
+export { IndexSessionError, resolveIndexSession };
+export type { IndexSession };
+
+/**
+ * Opens the replica for one operation and closes it afterwards.
+ *
+ * The key is fetched from the main process over the inherited fd for the duration
+ * of the call only and zeroed after (`withIndexKey`) - there is no cached handle
+ * and no resident key. A keychain round trip costs microseconds against work that
+ * makes network calls.
+ */
+export async function withReplica<T>(
+  accountId: string,
+  fn: (store: ReplicaStore) => Promise<T> | T,
+): Promise<T> {
+  const storeDir = getStoreDir();
+  if (!storeDir) {
+    throw new IndexSessionError('The offline replica is not enabled in this deployment.', 404);
+  }
+  return withIndexKey(accountId, async (key) => {
+    const store = ReplicaStore.open({ storeDir, accountId, key });
+    try {
+      return await fn(store);
+    } finally {
+      store.close();
+    }
+  });
+}
+
+/**
+ * The JMAP account whose mail is replicated.
+ *
+ * v1 replicates the PRIMARY mail account only. Every primary key already carries
+ * `jmap_account_id`, so adding the delegated/shared accounts a single login also
+ * exposes is inserting rows rather than a migration - JMAP ids are unique only
+ * WITHIN an account, and a schema that merged them would be cross-account leakage
+ * that costs nothing to prevent today and is unfixable later.
+ */
+export function primaryMailAccountId(session: JmapSessionInfo): string | null {
+  return accountIdFor(session, CAP_MAIL);
+}
+
+/**
+ * Single-flight per local account.
+ *
+ * On `globalThis` rather than in module scope for the same reason
+ * `lib/mail-index/key.ts` keeps its channel there: Next re-evaluates route
+ * modules (dev HMR, and separate module instances across route bundles), so a
+ * module-scoped map is not once-per-process and two overlapping requests would
+ * each get their own "single" flight. A Symbol key on globalThis is the one place
+ * in a Node process that survives module re-evaluation.
+ */
+const FLIGHT_KEY = Symbol.for('vncmail.offlineReplica.inFlight');
+
+function flights(): Map<string, Promise<CycleReport>> {
+  const holder = globalThis as unknown as Record<symbol, Map<string, Promise<CycleReport>> | undefined>;
+  const existing = holder[FLIGHT_KEY];
+  if (existing) return existing;
+  const created = new Map<string, Promise<CycleReport>>();
+  holder[FLIGHT_KEY] = created;
+  return created;
+}
+
+export interface SyncOptions {
+  /** Overrides the persisted policy for this cycle, and persists the override. */
+  policy?: RetentionPolicy;
+  /** Forces a rebuild: sets the sticky resync flag before the cycle runs. */
+  forceResync?: boolean;
+}
+
+/**
+ * Runs one cycle for the calling session's account, coalescing concurrent callers
+ * onto the same promise.
+ *
+ * Coalescing rather than aborting is deliberate: an implementation that set an
+ * abort flag and returned produced a cancelled sync and no new one - a "Sync now"
+ * tap during a sync did nothing at all. The in-flight promise is assigned to the
+ * map BEFORE the cycle body runs, because several early-return paths resolve
+ * synchronously and a later assignment leaves a re-entrancy hole; the cleanup is
+ * identity-checked so a slow loser cannot delete a newer flight.
+ */
+export async function syncAccount(
+  indexSession: IndexSession,
+  options: SyncOptions = {},
+): Promise<CycleReport> {
+  const map = flights();
+  const existing = map.get(indexSession.accountId);
+  if (existing) return existing;
+
+  const run = (async (): Promise<CycleReport> => {
+    // The session fetch is the FIRST network call of a cycle, so when the backend
+    // is unreachable this is where it fails - and it must be classified by the same
+    // taxonomy as everything else. Found by execution: without this, an offline
+    // sync surfaced a bare `JmapIndexError` 502 with no error class, so a caller
+    // could not tell "the network is down, retry later" from "this deployment is
+    // broken". "Offline is not an error" has to hold at the very first call too.
+    let session: JmapSessionInfo;
+    try {
+      session = await fetchJmapSession(indexSession.serverUrl, indexSession.authHeader);
+    } catch (error) {
+      const status = (error as { status?: number } | null)?.status;
+      const message = error instanceof Error ? error.message : String(error);
+      // `JmapIndexError.status` is OUR OWN value, not a server response status:
+      // it is 401 for auth, 429 for rate limiting, 504 for a timeout, and 502 for
+      // everything else - INCLUDING a `fetch` rejection with no server involved at
+      // all. So a bare 502 must be classified from the message, or "the machine is
+      // offline" is misread as "the server returned a 5xx". Passing the synthetic
+      // status straight into `classify` produced exactly that, found by the
+      // network-cut integration run.
+      if (status === 401 || status === 403) throw error;
+      const cls =
+        status === 429 ? 'RateLimit' as const
+        : status === 504 ? 'Transport' as const
+        : classify({ message });
+      throw new ReplicaSyncError(cls, message);
+    }
+    const jmapAccountId = primaryMailAccountId(session);
+    if (!jmapAccountId) {
+      throw new IndexSessionError('This account has no JMAP mail capability.', 409);
+    }
+
+    return withReplica(indexSession.accountId, async (store) => {
+      const now = Date.now();
+      if (options.policy) store.transaction(() => { store.setPolicy(options.policy as RetentionPolicy); });
+      if (options.forceResync) {
+        store.transaction(() => { store.patchFlags(now, { resyncRequired: true }); });
+      }
+      const policy = store.getPolicy();
+      const report = await runCycle({
+        store,
+        session,
+        authHeader: indexSession.authHeader,
+        jmapAccountId,
+        policy,
+        now,
+        deadline: now + BUDGET.wallClockMs,
+      });
+      logger.info('offline-replica: cycle complete', {
+        slot: indexSession.slot,
+        ok: report.ok,
+        phase: report.coveragePhase,
+        envelopes: report.envelopesWritten,
+        bodies: report.bodiesWritten,
+        deleted: report.envelopesDeleted,
+        unfinished: report.unfinishedWork,
+        warnings: report.warnings.length,
+        durationMs: report.durationMs,
+      });
+      return report;
+    });
+  })();
+
+  map.set(indexSession.accountId, run);
+  try {
+    return await run;
+  } finally {
+    if (map.get(indexSession.accountId) === run) map.delete(indexSession.accountId);
+  }
+}
+
+/**
+ * Resolves the JMAP account id for a READ without any network call.
+ *
+ * The read path must work with the backend unreachable, so it cannot fetch a JMAP
+ * session to learn the primary account id - that fetch is exactly what fails when
+ * offline. The store knows which account ids it holds rows for; with one
+ * replicated account that is unambiguous, and the caller may also pass an explicit
+ * id.
+ */
+export function resolveReadAccountId(store: ReplicaStore, requested?: string | null): string | null {
+  const known = store.knownJmapAccountIds();
+  if (requested && known.includes(requested)) return requested;
+  if (known.length === 1) return known[0];
+  if (requested) return null;
+  return known[0] ?? null;
+}
diff --git a/lib/offline-replica/errors.ts b/lib/offline-replica/errors.ts
new file mode 100644
index 00000000..bd976409
--- /dev/null
+++ b/lib/offline-replica/errors.ts
@@ -0,0 +1,151 @@
+// The error taxonomy. The whole point of this file is one column of one table:
+// **exactly one error class moves the cursor**, and its action is a full
+// verified rebuild. Everywhere else, failure means the cursor stands still.
+//
+// That is what makes "a failure never causes silent data loss" structural rather
+// than aspirational - and it is precisely what the mobile client's shipped
+// defect D4 got wrong, by collapsing every error to `null` and then adopting a
+// snapshot state as the next cursor. A transient 503 on `Email/changes` was
+// enough to fast-forward the cursor over every change the client had not seen.
+
+export type ErrorClass =
+  | 'Transport'
+  | 'RateLimit'
+  | 'ServerTransient'
+  | 'RequestLimit'
+  | 'Auth'
+  | 'Fatal'
+  | 'StateInvalid';
+
+/** True for the one class that moves a cursor - and it moves it to "invalidated". */
+export function movesCursor(cls: ErrorClass): boolean {
+  return cls === 'StateInvalid';
+}
+
+/**
+ * Only a size/availability problem is worth escalating to a rebuild.
+ *
+ * Escalating on RateLimit would mean the response to a rate-limited server is to
+ * issue far MORE requests - a full window re-enumeration. Escalating on Auth
+ * would let a 401 trigger a rebuild; on Transport, a flaky tunnel would do the
+ * same. Fatal is our own bug and a rebuild will not fix it.
+ */
+export function escalationApplies(cls: ErrorClass): boolean {
+  return cls === 'ServerTransient' || cls === 'RequestLimit';
+}
+
+/** JMAP method-level error types that invalidate a `/changes` cursor. */const STATE_INVALID_TYPES = new Set(['cannotCalculateChanges']);
+
+const FATAL_TYPES = new Set([
+  'invalidArguments', 'unknownMethod', 'accountNotFound', 'forbidden',
+  'unsupportedFilter', 'unsupportedSort', 'invalidResultReference',
+  'accountNotSupportedByMethod', 'accountReadOnly',
+]);
+
+const REQUEST_LIMIT_TYPES = new Set([
+  'maxSizeRequest', 'maxCallsInRequest', 'requestTooLarge', 'maxObjectsInGet',
+  'tooLarge',
+]);
+
+const SERVER_TRANSIENT_TYPES = new Set([
+  'serverUnavailable', 'serverFail', 'serverPartialFail', 'stateMismatch',
+]);
+
+export class ReplicaSyncError extends Error {
+  readonly cls: ErrorClass;
+  readonly retryAfterMs?: number;
+  constructor(cls: ErrorClass, message: string, retryAfterMs?: number) {
+    super(message);
+    this.name = 'ReplicaSyncError';
+    this.cls = cls;
+    this.retryAfterMs = retryAfterMs;
+  }
+}
+
+/**
+ * Classification is STRUCTURE BEFORE STRINGS: HTTP status, then JMAP error type,
+ * and only then message prose. A method error's `description` can legitimately
+ * contain the words "timeout" or "socket", and `fetch failed: ECONNRESET` must
+ * not be read as a JMAP method error.
+ */
+export function classify(input: {
+  httpStatus?: number;
+  jmapErrorType?: string;
+  message?: string;
+}): ErrorClass {
+  const { httpStatus, jmapErrorType, message } = input;
+
+  if (typeof httpStatus === 'number') {
+    if (httpStatus === 401 || httpStatus === 403) return 'Auth';
+    if (httpStatus === 429) return 'RateLimit';
+    if (httpStatus === 413) return 'RequestLimit';
+    if (httpStatus >= 500) return 'ServerTransient';
+  }
+
+  if (jmapErrorType) {
+    if (STATE_INVALID_TYPES.has(jmapErrorType)) return 'StateInvalid';
+    if (REQUEST_LIMIT_TYPES.has(jmapErrorType)) return 'RequestLimit';
+    if (FATAL_TYPES.has(jmapErrorType)) return 'Fatal';
+    if (SERVER_TRANSIENT_TYPES.has(jmapErrorType)) return 'ServerTransient';
+    if (jmapErrorType === 'limit') return 'RateLimit';
+    // An UNRECOGNISED method-level error is ServerTransient, never Fatal and
+    // never StateInvalid. Guessing transient costs a retry; guessing
+    // state-invalid costs a full resync; guessing fatal stalls the account. The
+    // cheapest wrong answer wins the default.
+    return 'ServerTransient';
+  }
+
+  if (message) {
+    const lower = message.toLowerCase();
+    if (
+      lower.includes('fetch failed') || lower.includes('econnrefused') ||
+      lower.includes('econnreset') || lower.includes('enotfound') ||
+      lower.includes('etimedout') || lower.includes('socket') ||
+      lower.includes('network') || lower.includes('timed out') ||
+      lower.includes('eai_again') || lower.includes('ehostunreach') ||
+      lower.includes('enetunreach') || lower.includes('certificate')
+    ) {
+      // "Offline is not an error." Transport failures leave every cursor exactly
+      // where it was and are retried later.
+      return 'Transport';
+    }
+  }
+
+  return 'ServerTransient';
+}
+
+/** Full-jitter exponential backoff. */
+export function backoffDelayMs(attempt: number, opts: { baseMs?: number; capMs?: number } = {}): number {
+  const base = opts.baseMs ?? 1_000;
+  const cap = opts.capMs ?? 60_000;
+  const ceiling = Math.min(cap, base * 2 ** Math.max(0, attempt));
+  // Jitter is not decoration: several triggers fire at once (launch catch-up,
+  // network recovery, a push burst) against one Stalwart instance, which is
+  // exactly the shape that produces a synchronised stampede.
+  return Math.floor(Math.random() * ceiling);
+}
+
+/**
+ * The `maxChanges` ladder, monotonically SHRINKING, every rung expressed
+ * relative to rung 0.
+ *
+ * Two bugs live here historically. First, an unbounded middle rung produced a
+ * retry strictly LARGER than the attempt that just failed - actively worsening a
+ * "response too large" error. Then clamping only rung 0 reintroduced it in a
+ * narrower form: a server advertising `maxObjectsInGet: 100` gave rung 0 = 100
+ * and rung 1 = 250. Deriving every rung from rung 0 is what makes
+ * monotonic-non-increase true for every server value.
+ */
+export function rungValue(rung: 0 | 1 | 2 | 3, maxObjectsInGet: number | undefined): number {
+  const rung0 = Math.max(1, Math.min(maxObjectsInGet ?? 500, 500));
+  switch (rung) {
+    case 0: return rung0;
+    case 1: return Math.max(1, Math.min(rung0, 250));
+    case 2: return Math.max(1, Math.min(rung0, 50));
+    case 3: return Math.max(1, Math.min(rung0, 25));
+  }
+}
+
+export function nextRung(rung: 0 | 1 | 2 | 3): 0 | 1 | 2 | 3 {
+  return rung >= 3 ? 3 : ((rung + 1) as 0 | 1 | 2 | 3);
+}
diff --git a/lib/offline-replica/jmap.ts b/lib/offline-replica/jmap.ts
new file mode 100644
index 00000000..874425f7
--- /dev/null
+++ b/lib/offline-replica/jmap.ts
@@ -0,0 +1,302 @@
+// The replica's JMAP calls. Reuses `lib/mail-index/jmap.ts`'s session fetch,
+// origin pinning and request plumbing rather than duplicating them (that file
+// already handles Stalwart's 307 on /.well-known/jmap and refuses to send
+// credentials off-origin), and adds the delta-sync methods the index never
+// needed: `Mailbox/changes`, `Email/changes`, the ascending coverage query, and
+// the two-tier `Email/get`.
+//
+// THIS FILE IS THE ONLY PLACE ALLOWED TO MINT A BRANDED STATE TOKEN. That is what
+// makes cursor provenance checkable by grep: `asChangesState` appears only in the
+// `/changes` parser, `asSnapshotState` only in the `Foo/get {ids: []}` parser.
+
+import type { Email, Mailbox } from '@/lib/jmap/types';
+import { CAP_CORE, CAP_MAIL, jmapRequest, type JmapSessionInfo } from '@/lib/mail-index/jmap';
+import type { ChangesPage } from './apply';
+import { ReplicaSyncError, classify } from './errors';
+import { asChangesState, asSnapshotState, type SnapshotState } from './states';
+
+/** The envelope tier. Mirrors `lib/jmap/client.ts`'s EMAIL_LIST_PROPERTIES exactly. */
+export const ENVELOPE_PROPERTIES = [
+  'id', 'threadId', 'mailboxIds', 'keywords', 'size', 'receivedAt',
+  'from', 'to', 'cc', 'subject', 'preview', 'hasAttachment', 'blobId',
+] as const;
+
+/**
+ * The body tier - everything `lib/jmap/client.ts`'s `getEmail()` asks for beyond
+ * the envelope tier, so a replica-served message is field-for-field what the
+ * online read path produces. `components/email/email-viewer.tsx` reads
+ * `bodyValues` keyed by the SAME partIds as `htmlBody`/`textBody`, so all three
+ * must travel together or the viewer sits on its loading skeleton forever.
+ */
+export const BODY_PROPERTIES = [
+  'id', 'sentAt', 'bcc', 'replyTo', 'textBody', 'htmlBody', 'bodyValues',
+  'attachments', 'messageId', 'inReplyTo', 'references', 'headers', 'bodyStructure',
+] as const;
+
+/** The three MUTABLE properties. */
+export const MUTABLE_PROPERTIES = ['id', 'keywords', 'mailboxIds'] as const;
+
+export const MAX_BODY_VALUE_BYTES = 512_000;
+
+interface MethodError {
+  type?: string;
+  description?: string;
+}
+
+function asMethodError(args: Record<string, unknown>): MethodError {
+  return {
+    type: typeof args.type === 'string' ? args.type : undefined,
+    description: typeof args.description === 'string' ? args.description : undefined,
+  };
+}
+
+/**
+ * Runs one JMAP request and classifies any failure. Wraps the shared transport so
+ * a transport-level failure becomes `Transport` (cursor untouched) rather than an
+ * opaque throw the caller has to guess about.
+ */
+async function call(
+  session: JmapSessionInfo,
+  authHeader: string,
+  methodCalls: ReadonlyArray<[string, Record<string, unknown>, string]>,
+): Promise<Array<[string, Record<string, unknown>, string]>> {
+  try {
+    return await jmapRequest(session, authHeader, [CAP_CORE, CAP_MAIL], methodCalls);
+  } catch (error) {
+    const status = (error as { status?: number } | null)?.status;
+    const message = error instanceof Error ? error.message : String(error);
+    throw new ReplicaSyncError(classify({ httpStatus: status, message }), message);
+  }
+}
+
+function findResponse(
+  responses: Array<[string, Record<string, unknown>, string]>,
+  callId: string,
+): { name: string; args: Record<string, unknown> } | null {
+  for (const [name, args, id] of responses) {
+    if (id === callId) return { name, args };
+  }
+  return null;
+}
+
+/** Turns a method-level `error` response into a classified throw. */
+function raiseMethodError(args: Record<string, unknown>, context: string): never {
+  const { type, description } = asMethodError(args);
+  throw new ReplicaSyncError(
+    classify({ jmapErrorType: type, message: description }),
+    `${context} failed: ${type ?? 'unknown'}${description ? ` (${description})` : ''}`,
+  );
+}
+
+function strArray(v: unknown): string[] {
+  return Array.isArray(v) ? v.filter((x): x is string => typeof x === 'string') : [];
+}
+
+// ── snapshot states, for bootstrap / reconcile ───────────────────────────────
+
+/**
+ * Captures both cursors in ONE request, before touching any data.
+ *
+ * `Foo/get {ids: []}` returns the account's current state token with no records,
+ * which RFC 8620 s5.1 defines as a valid `sinceState` for `Foo/changes`. This is
+ * step 1 of the mandatory bootstrap order and the single thing most likely to be
+ * "optimised" into a permanent data hole: the cursor must be captured BEFORE the
+ * enumeration, so it is deliberately OLDER than the data and the first delta cycle
+ * re-delivers a few changes we already have. The cheaper opposite order - enumerate,
+ * then capture - silently loses every change that arrived during the scan, which on
+ * a large mailbox is minutes.
+ */
+export async function captureSnapshotStates(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+): Promise<{ mailbox: SnapshotState; email: SnapshotState }> {
+  const responses = await call(session, authHeader, [
+    ['Mailbox/get', { accountId, ids: [] }, 'm'],
+    ['Email/get', { accountId, ids: [] }, 'e'],
+  ]);
+  const mailbox = findResponse(responses, 'm');
+  const email = findResponse(responses, 'e');
+  if (!mailbox || mailbox.name === 'error') {
+    raiseMethodError(mailbox?.args ?? {}, 'Mailbox/get (state capture)');
+  }
+  if (!email || email.name === 'error') {
+    raiseMethodError(email?.args ?? {}, 'Email/get (state capture)');
+  }
+  return {
+    mailbox: asSnapshotState(mailbox.args.state),
+    email: asSnapshotState(email.args.state),
+  };
+}
+
+// ── /changes ─────────────────────────────────────────────────────────────────
+
+function parseChangesPage(args: Record<string, unknown>): ChangesPage {
+  return {
+    oldState: asChangesState(args.oldState),
+    newState: asChangesState(args.newState),
+    hasMoreChanges: args.hasMoreChanges === true,
+    created: strArray(args.created),
+    updated: strArray(args.updated),
+    destroyed: strArray(args.destroyed),
+    updatedProperties:
+      args.updatedProperties === null
+        ? null
+        : Array.isArray(args.updatedProperties)
+          ? strArray(args.updatedProperties)
+          : undefined,
+  };
+}
+
+export async function getMailboxChanges(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+  sinceState: string,
+  maxChanges: number,
+): Promise<ChangesPage> {
+  const responses = await call(session, authHeader, [
+    ['Mailbox/changes', { accountId, sinceState, maxChanges }, 'c'],
+  ]);
+  const res = findResponse(responses, 'c');
+  if (!res || res.name === 'error') raiseMethodError(res?.args ?? {}, 'Mailbox/changes');
+  return parseChangesPage(res.args);
+}
+
+export async function getEmailChanges(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+  sinceState: string,
+  maxChanges: number,
+): Promise<ChangesPage> {
+  const responses = await call(session, authHeader, [
+    ['Email/changes', { accountId, sinceState, maxChanges }, 'c'],
+  ]);
+  const res = findResponse(responses, 'c');
+  if (!res || res.name === 'error') raiseMethodError(res?.args ?? {}, 'Email/changes');
+  // `Email/changes` has no `updatedProperties` - RFC 8621 s4.3 is a plain
+  // /changes - which is why the 3-property `Email/get` is unavoidable there.
+  return parseChangesPage(res.args);
+}
+
+// ── gets ─────────────────────────────────────────────────────────────────────
+
+export async function getMailboxes(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+  ids: readonly string[] | null,
+  properties?: readonly string[],
+): Promise<Mailbox[]> {
+  const args: Record<string, unknown> = { accountId, ids: ids === null ? null : [...ids] };
+  if (properties) args.properties = [...properties, 'id'];
+  const responses = await call(session, authHeader, [['Mailbox/get', args, 'g']]);
+  const res = findResponse(responses, 'g');
+  if (!res || res.name === 'error') raiseMethodError(res?.args ?? {}, 'Mailbox/get');
+  return Array.isArray(res.args.list) ? (res.args.list as Mailbox[]) : [];
+}
+
+export interface EmailGetResult {
+  list: Email[];
+  /** Normal, not an error: the record was destroyed between /changes and /get. */
+  notFound: string[];
+}
+
+export async function getEmails(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+  ids: readonly string[],
+  tier: 'envelope' | 'mutable' | 'body',
+): Promise<EmailGetResult> {
+  if (ids.length === 0) return { list: [], notFound: [] };
+  const args: Record<string, unknown> = { accountId, ids: [...ids] };
+  if (tier === 'envelope') {
+    args.properties = [...ENVELOPE_PROPERTIES];
+  } else if (tier === 'mutable') {
+    args.properties = [...MUTABLE_PROPERTIES];
+  } else {
+    args.properties = [...BODY_PROPERTIES];
+    // Without these the bodyValues map comes back EMPTY and every stored body
+    // would be an empty object that renders as a blank message offline.
+    args.fetchTextBodyValues = true;
+    args.fetchHTMLBodyValues = true;
+    args.fetchAllBodyValues = true;
+    args.maxBodyValueBytes = MAX_BODY_VALUE_BYTES;
+  }
+  const responses = await call(session, authHeader, [['Email/get', args, 'g']]);
+  const res = findResponse(responses, 'g');
+  if (!res || res.name === 'error') raiseMethodError(res?.args ?? {}, 'Email/get');
+  return {
+    list: Array.isArray(res.args.list) ? (res.args.list as Email[]) : [],
+    notFound: strArray(res.args.notFound),
+  };
+}
+
+// ── coverage enumeration ─────────────────────────────────────────────────────
+
+export interface CoveragePage {
+  ids: string[];
+  /** Echoed back so the caller can detect the tie-cluster case. */
+  requestedAfter: string;
+}
+
+/**
+ * The ascending keyset walk.
+ *
+ * ASCENDING is not a style choice. New mail arrives at the TAIL, so insertions
+ * never shift rows the scan has already passed. With a DESCENDING sort and
+ * position-based paging, one delivery between page 1 and page 2 pushes a message
+ * from page 1's boundary into page 2's start and one message out of the scan's
+ * reach entirely - and that message is pre-existing relative to our cursor, so
+ * `Email/changes` will never report it. A permanent hole with no signal it exists.
+ *
+ * `calculateTotal: false` because the total is unstable and unused.
+ */
+export async function queryAscending(
+  session: JmapSessionInfo,
+  authHeader: string,
+  accountId: string,
+  after: string,
+  limit: number,
+  anchor?: { anchor: string; anchorOffset: number },
+): Promise<CoveragePage> {
+  const args: Record<string, unknown> = {
+    accountId,
+    filter: { after },
+    sort: [{ property: 'receivedAt', isAscending: true }],
+    limit,
+    calculateTotal: false,
+  };
+  if (anchor) {
+    args.anchor = anchor.anchor;
+    args.anchorOffset = anchor.anchorOffset;
+  }
+  const responses = await call(session, authHeader, [['Email/query', args, 'q']]);
+  const res = findResponse(responses, 'q');
+  if (!res || res.name === 'error') {
+    const { type } = asMethodError(res?.args ?? {});
+    if (type === 'anchorNotFound') {
+      // Not a failure - the caller falls back to the last-resort rung.
+      throw new AnchorNotFoundError();
+    }
+    raiseMethodError(res?.args ?? {}, 'Email/query');
+  }
+  return { ids: strArray(res.args.ids), requestedAfter: after };
+}
+
+export class AnchorNotFoundError extends Error {
+  constructor() {
+    super('Email/query rejected the anchor');
+    this.name = 'AnchorNotFoundError';
+  }
+}
+
+/** `maxObjectsInGet`, so the maxChanges ladder can be clamped to what the server allows. */
+export function maxObjectsInGet(session: JmapSessionInfo): number | undefined {
+  const core = session.capabilities[CAP_CORE];
+  if (!core || typeof core !== 'object') return undefined;
+  const value = (core as Record<string, unknown>).maxObjectsInGet;
+  return typeof value === 'number' && Number.isFinite(value) ? value : undefined;
+}
diff --git a/lib/offline-replica/read.ts b/lib/offline-replica/read.ts
new file mode 100644
index 00000000..c0248d56
--- /dev/null
+++ b/lib/offline-replica/read.ts
@@ -0,0 +1,219 @@
+// The offline READ path: stored rows back into the exact `Email` / `Mailbox`
+// shapes `lib/jmap/client.ts` returns, so the renderer cannot tell the difference.
+//
+// COHERENCE (the review's H3, the one finding that genuinely returns once a
+// replica exists). The webmail already does LOCAL DELTA ARITHMETIC on mailbox
+// unread counts and totals for mark-read/move/delete, with a comment referencing
+// a production bug from getting that cutoff wrong. A read-only cache sitting
+// underneath that arithmetic needs an explicit coherence story, and the story is:
+//
+//   THE REPLICA IS A FALLBACK, NEVER A CACHE IN FRONT OF THE SERVER.
+//
+// It is consulted only after a read has actually failed at the transport level
+// (see `lib/offline-fallback-client.ts`), so an online session never sees a
+// replica count and the arithmetic never operates on replica numbers. While
+// offline, counts are whatever the last successful sync recorded and any local
+// mark-read drift is bounded, invisible in the same session, and repaired by the
+// next `Mailbox/changes` - which is the authoritative source for all four
+// counters. The alternative - serving the replica first and reconciling - is what
+// would need the coherence rules the review asked for, and is not what this does.
+//
+// Every shape here is intentionally what the ONLINE path produces, including
+// `parseEmailHeaders`' derived security fields, because
+// `components/email/email-viewer.tsx` reads them directly. In particular
+// `bodyValues` must be keyed by the same partIds as `htmlBody`/`textBody`, or the
+// viewer's `isBodyLoading` gate sits on its skeleton forever.
+
+import { parseAuthenticationResults, parseSpamLLM, parseSpamScore } from '@/lib/email-headers';
+import type { Email, EmailAddress, Mailbox } from '@/lib/jmap/types';
+import type { ReplicaStore } from './store';
+import type { MailboxRow } from './types';
+
+const DEFAULT_RIGHTS: Mailbox['myRights'] = {
+  mayReadItems: true, mayAddItems: false, mayRemoveItems: false, maySetSeen: false,
+  maySetKeywords: false, mayCreateChild: false, mayRename: false, mayDelete: false,
+  maySubmit: false,
+};
+
+function parseJson<T>(raw: unknown, fallback: T): T {
+  if (typeof raw !== 'string' || raw.length === 0) return fallback;
+  try {
+    return JSON.parse(raw) as T;
+  } catch {
+    return fallback;
+  }
+}
+
+export function rowToMailbox(row: MailboxRow): Mailbox {
+  return {
+    id: row.id,
+    name: row.name,
+    parentId: row.parentId ?? undefined,
+    role: row.role ?? undefined,
+    sortOrder: row.sortOrder ?? 0,
+    totalEmails: row.totalEmails ?? 0,
+    unreadEmails: row.unreadEmails ?? 0,
+    totalThreads: row.totalThreads ?? 0,
+    unreadThreads: row.unreadThreads ?? 0,
+    // Offline, the rights that matter are the read ones. Every mutating right
+    // defaults to false so no UI offers an action that cannot possibly succeed
+    // with no network; the real rights return with the next sync.
+    myRights: parseJson<Mailbox['myRights']>(row.myRightsJson, DEFAULT_RIGHTS),
+    isSubscribed: row.isSubscribed,
+  };
+}
+
+/** The envelope tier, as `getEmails()` would return it. */
+export function rowToEnvelope(row: Record<string, unknown>, mailboxIds: readonly string[]): Email {
+  const keywords = parseJson<Record<string, boolean>>(row.keywords_json, {});
+  const mailboxMap: Record<string, boolean> = {};
+  for (const id of mailboxIds) mailboxMap[id] = true;
+  return {
+    id: String(row.id),
+    threadId: typeof row.thread_id === 'string' ? row.thread_id : String(row.id),
+    mailboxIds: mailboxMap,
+    keywords,
+    size: typeof row.size === 'number' ? row.size : 0,
+    receivedAt: String(row.received_at),
+    from: parseJson<EmailAddress[] | undefined>(row.from_json, undefined),
+    to: parseJson<EmailAddress[] | undefined>(row.to_json, undefined),
+    cc: parseJson<EmailAddress[] | undefined>(row.cc_json, undefined),
+    subject: typeof row.subject === 'string' ? row.subject : undefined,
+    preview: typeof row.preview === 'string' ? row.preview : undefined,
+    hasAttachment: row.has_attachment !== 0,
+    blobId: typeof row.blob_id === 'string' ? row.blob_id : undefined,
+  };
+}
+
+interface StoredBody {
+  sentAt?: string;
+  bcc?: EmailAddress[];
+  replyTo?: EmailAddress[];
+  textBody?: Email['textBody'];
+  htmlBody?: Email['htmlBody'];
+  bodyValues?: Email['bodyValues'];
+  attachments?: Email['attachments'];
+  messageId?: string;
+  inReplyTo?: string[];
+  references?: string[];
+  headers?: unknown;
+  bodyStructure?: Email['bodyStructure'];
+}
+
+/**
+ * Normalises JMAP's `headers` array into the record shape the renderer expects
+ * and derives the security fields, reproducing what `JMAPClient`'s private
+ * `parseEmailHeaders` does on the online path.
+ *
+ * Reproduced here rather than imported because `lib/jmap/client.ts` is a
+ * 7400-line renderer object that holds credentials in instance fields, opens push
+ * connections and wires itself into Zustand stores - importing it into a server
+ * route would drag all of that into the server bundle. The parsing HELPERS in
+ * `lib/email-headers` are shared, so the only duplicated logic is the array->record
+ * flattening.
+ */
+function applyHeaders(email: Email, rawHeaders: unknown): void {
+  let record: Record<string, string | string[]>;
+  if (Array.isArray(rawHeaders)) {
+    record = {};
+    for (const header of rawHeaders as Array<{ name?: string; value?: string }>) {
+      if (!header?.name || !header?.value) continue;
+      const existing = record[header.name];
+      if (existing) {
+        record[header.name] = Array.isArray(existing)
+          ? [...existing, header.value]
+          : [existing, header.value];
+      } else {
+        record[header.name] = header.value;
+      }
+    }
+  } else if (rawHeaders && typeof rawHeaders === 'object') {
+    record = rawHeaders as Record<string, string | string[]>;
+  } else {
+    return;
+  }
+  email.headers = record;
+
+  const authResults = record['Authentication-Results'];
+  if (authResults) {
+    const value = Array.isArray(authResults) ? authResults.join('; ') : authResults;
+    email.authenticationResults = parseAuthenticationResults(value);
+  }
+
+  for (const name of ['X-Spam-Score', 'X-Spam-Status', 'X-Spam-Result', 'X-Rspamd-Score']) {
+    const header = record[name];
+    if (!header) continue;
+    const value = Array.isArray(header) ? header[0] : header;
+    const parsed = parseSpamScore(String(value).trim());
+    if (parsed) {
+      email.spamScore = parsed.score;
+      email.spamStatus = parsed.status;
+      break;
+    }
+  }
+
+  const llm = record['X-Spam-LLM'];
+  if (llm) {
+    const parsed = parseSpamLLM(String(Array.isArray(llm) ? llm[0] : llm));
+    if (parsed) email.spamLLM = parsed;
+  }
+}
+
+export interface OfflineMessage {
+  email: Email;
+  /** False when only the envelope is held, so the caller can say so rather than render blank. */
+  hasBody: boolean;
+}
+
+/** One full message, envelope + body, exactly as `getEmail()` would return it. */
+export function readMessage(
+  store: ReplicaStore,
+  jmapAccountId: string,
+  id: string,
+): OfflineMessage | null {
+  const row = store.getEnvelopeRaw(jmapAccountId, id);
+  if (!row) return null;
+  const email = rowToEnvelope(row, store.mailboxIdsFor(jmapAccountId, id));
+
+  const bodyJson = store.getBody(jmapAccountId, id);
+  if (!bodyJson) return { email, hasBody: false };
+
+  const body = parseJson<StoredBody>(bodyJson, {});
+  email.sentAt = body.sentAt;
+  email.bcc = body.bcc;
+  email.replyTo = body.replyTo;
+  email.textBody = body.textBody;
+  email.htmlBody = body.htmlBody;
+  email.bodyValues = body.bodyValues;
+  email.attachments = body.attachments;
+  email.messageId = body.messageId;
+  email.inReplyTo = body.inReplyTo;
+  email.references = body.references;
+  email.bodyStructure = body.bodyStructure;
+  applyHeaders(email, body.headers);
+
+  // A body row whose `bodyValues` came back empty would render as a blank
+  // message and, worse, leave the viewer's loading gate stuck. Report it as
+  // "envelope only" instead, which the UI can explain.
+  const hasBody =
+    !!email.bodyValues && Object.keys(email.bodyValues).length > 0;
+  return { email, hasBody };
+}
+
+export function readMailboxes(store: ReplicaStore, jmapAccountId: string): Mailbox[] {
+  return store.listMailboxes(jmapAccountId).map(rowToMailbox);
+}
+
+export function readEnvelopePage(
+  store: ReplicaStore,
+  jmapAccountId: string,
+  mailboxId: string | null,
+  limit: number,
+  offset: number,
+): { emails: Email[]; total: number; hasMore: boolean } {
+  const { rows, total } = store.listEnvelopes(jmapAccountId, mailboxId, limit, offset);
+  const emails = rows.map((row) =>
+    rowToEnvelope(row, store.mailboxIdsFor(jmapAccountId, String(row.id))),
+  );
+  return { emails, total, hasMore: offset + emails.length < total };
+}
diff --git a/lib/offline-replica/retention.ts b/lib/offline-replica/retention.ts
new file mode 100644
index 00000000..77636cc6
--- /dev/null
+++ b/lib/offline-replica/retention.ts
@@ -0,0 +1,149 @@
+// Retention floors, and the clock-jump guard.
+//
+// THE BUG THIS FILE IS SHAPED BY. No cursor and no ordering in this engine
+// depends on the device clock - but the retention BOUNDARY does, so a large clock
+// skew moves the window. The mobile client added a guard: if the computed floor
+// moves more than ~25 h from the last one, keep the previous floor and warn.
+//
+// The guard then wiped the entire offline store. Mechanism, exactly:
+//
+//   1. clock jumps forward a year
+//   2. cycle N computes a floor a year ahead, detects the >25 h move, holds the
+//      old floor for this cycle - and PERSISTS THE COMPUTED (jumped) FLOOR as
+//      `lastWindowFloor`
+//   3. the follow-on cycle a few seconds later computes a floor within seconds of
+//      the persisted one, so `delta <= threshold`, so the guard passes
+//   4. that floor is classified as a retention NARROW, and every envelope below
+//      it is evicted - i.e. all of them, since the floor is a year in the future
+//   5. `coveredFrom` then claims the range complete, and `/changes` cannot
+//      re-deliver pre-existing mail. Unrecoverable.
+//
+// A clock glitch wiped the store about five seconds after being detected,
+// THROUGH the very mechanism meant to prevent that. Its reproduction returned 0
+// envelopes from a 4-envelope store.
+//
+// The generalisable lesson, worth more than the code: a guard that DETECTS an
+// anomaly but PERSISTS the anomalous value converts a transient glitch into a
+// legitimised new baseline. Any "suppress and remember" guard must remember the
+// value it USED, not the value it rejected - and must expose a separate
+// "don't delete anything on this basis" bit, because suppressing the floor is not
+// the same as suppressing the deletions the floor authorises.
+
+import type { RetentionPolicy } from './store';
+
+/** A day plus an hour: an ordinary DST shift or NTP correction must not trip it. */
+export const CLOCK_JUMP_GUARD_MS = 25 * 60 * 60 * 1000;
+
+export interface RetentionFloors {
+  envelopeFrom: string;
+  bodyFrom: string;
+  maxBodyBytes: number;
+}
+
+function isoDaysAgo(now: number, days: number): string {
+  return new Date(now - days * 24 * 60 * 60 * 1000).toISOString();
+}
+
+export function computeFloors(policy: RetentionPolicy, now: number): RetentionFloors {
+  // The body window can never be wider than the envelope window: a body with no
+  // envelope is an orphan by construction, and the whole point of two tiers is
+  // envelopes being a superset of bodies.
+  const bodyDays = Math.min(policy.bodyDays, policy.envelopeDays);
+  return {
+    envelopeFrom: isoDaysAgo(now, policy.envelopeDays),
+    bodyFrom: isoDaysAgo(now, bodyDays),
+    maxBodyBytes: Math.max(0, Math.floor(policy.maxBodyMB * 1024 * 1024)),
+  };
+}
+
+export interface GuardedFloor {
+  /** The floor to actually use for eviction and for any sweep. */
+  envelopeFrom: string;
+  /** True when the guard suppressed a suspicious jump. */
+  suppressed: boolean;
+  /**
+   * False while the floor in use came from a SUSPECT clock reading. Retention
+   * eviction and the reconcile sweep must BOTH refuse to act on it - suppressing
+   * the floor is not the same as suppressing the deletions it authorises.
+   */
+  evictionAllowed: boolean;
+  /** What to persist as `lastWindowFloor`. */
+  nextLastWindowFloor: string;
+  warning?: string;
+}
+
+export function guardFloorAgainstClockJump(
+  computed: string,
+  lastWindowFloor: string | undefined,
+  opts: { policyChanged?: boolean } = {},
+): GuardedFloor {
+  const adopt = (floor: string): GuardedFloor => ({
+    envelopeFrom: floor,
+    suppressed: false,
+    evictionAllowed: true,
+    nextLastWindowFloor: floor,
+  });
+
+  if (!lastWindowFloor) return adopt(computed);
+
+  // An explicit `envelopeDays` change is INTENT, not a glitch, and must take
+  // effect - including its eviction. This is also why "adopt on the second
+  // consistent observation" had to go: with intent handled here, that rule
+  // existed ONLY for the clock-anomaly case, i.e. only for the case where
+  // adopting is the harmful thing to do.
+  if (opts.policyChanged) return adopt(computed);
+
+  const delta = Math.abs(Date.parse(computed) - Date.parse(lastWindowFloor));
+  if (!Number.isFinite(delta) || delta <= CLOCK_JUMP_GUARD_MS) return adopt(computed);
+
+  // SUSPECT. Ignore the reading entirely, keep the previous floor, and persist
+  // THE PREVIOUS FLOOR - so every subsequent cycle re-detects the same jump and
+  // stays suppressed. Persisting the computed value here is the wipe described
+  // at the top of this file.
+  //
+  // The trade-off, stated rather than hidden: on a device whose clock is
+  // permanently wrong by more than a day, retention stops tracking the clock and
+  // the store keeps MORE mail than the setting says. That is bounded (envelopes
+  // are ~1 KB, bodies are capped in bytes) and self-clears the moment the user
+  // changes a retention setting or the clock returns. Keeping too much mail is
+  // the correct direction to fail for a feature whose entire purpose is having
+  // mail available offline.
+  return {
+    envelopeFrom: lastWindowFloor,
+    suppressed: true,
+    evictionAllowed: false,
+    nextLastWindowFloor: lastWindowFloor,
+    warning:
+      `retention floor moved ${Math.round(delta / 3_600_000)}h in one step ` +
+      `(${lastWindowFloor} -> ${computed}); treating it as a clock anomaly and ` +
+      `refusing to evict or sweep on this basis`,
+  };
+}
+
+export type FloorMovement = 'unchanged' | 'widened' | 'narrowed';
+
+export function floorMovement(previous: string | undefined, next: string): FloorMovement {
+  if (!previous) return 'unchanged';
+  if (next === previous) return 'unchanged';
+  // A LATER floor keeps less mail.
+  return next > previous ? 'narrowed' : 'widened';
+}
+
+export interface WindowAdjustment {
+  /** Set when envelopes below this floor should be evicted. */
+  evictBelow?: string;
+  /** Set when coverage must re-scan from a wider floor. */
+  rescanFrom?: string;
+}
+
+/**
+ * What a floor movement asks for.
+ *
+ * A WIDEN is not a resync: `targetFrom` moves back, coverage re-enters
+ * `scanning`, and the cursors are untouched. A NARROW evicts.
+ */
+export function adjustForWindow(movement: FloorMovement, floor: string): WindowAdjustment {
+  if (movement === 'narrowed') return { evictBelow: floor };
+  if (movement === 'widened') return { rescanFrom: floor };
+  return {};
+}
diff --git a/lib/offline-replica/route-gate.ts b/lib/offline-replica/route-gate.ts
new file mode 100644
index 00000000..dcb734b2
--- /dev/null
+++ b/lib/offline-replica/route-gate.ts
@@ -0,0 +1,67 @@
+// The gate every replica route shares, and its error mapping.
+//
+// 404, not 403, when the feature is absent: the standalone server artifact is the
+// SAME one the production Dockerfile ships to multi-tenant deployments, where a
+// server-side replica of every user's mail would be badly wrong. Nothing should
+// learn the routes exist in a deployment that does not have the feature.
+
+import { NextResponse } from 'next/server';
+import { logger } from '@/lib/logger';
+import { isSqlcipherAvailable } from '@/lib/mail-index/binding';
+import { hasKeyChannel, IndexKeyError } from '@/lib/mail-index/key';
+import { getStoreDir } from '@/lib/mail-index/paths';
+import { JmapIndexError } from '@/lib/mail-index/jmap';
+import { IndexSessionError } from '@/lib/mail-index/reindex';
+import { ReplicaUnavailableError } from './store';
+import { ReplicaSyncError } from './errors';
+
+/** Returns a response to send immediately, or `null` when the gate is open. */
+export function gateReplicaRoute(): NextResponse | null {
+  if (!getStoreDir()) return new NextResponse(null, { status: 404 });
+  if (!hasKeyChannel()) {
+    return NextResponse.json(
+      { error: 'The offline replica has no key channel in this process.', code: 'no-key-channel' },
+      { status: 503 },
+    );
+  }
+  if (!isSqlcipherAvailable()) {
+    // The native binding is an optionalDependency, so "not installed" is a normal
+    // state on a platform with no prebuild - not an error to log loudly.
+    return NextResponse.json(
+      { error: 'Encrypted local storage is unavailable on this platform.', code: 'no-binding' },
+      { status: 503 },
+    );
+  }
+  return null;
+}
+
+export function replicaErrorResponse(error: unknown, context: string): NextResponse {
+  if (error instanceof IndexSessionError) {
+    return NextResponse.json({ error: error.message }, { status: error.status });
+  }
+  if (error instanceof JmapIndexError) {
+    return NextResponse.json({ error: error.message }, { status: error.status });
+  }
+  if (error instanceof IndexKeyError) {
+    // `no-secure-storage` is the Linux-without-a-keyring refusal: a real, expected
+    // outcome with a user-facing explanation, not a server fault.
+    const status = error.code === 'no-secure-storage' ? 503 : 500;
+    return NextResponse.json({ error: error.message, code: error.code }, { status });
+  }
+  if (error instanceof ReplicaUnavailableError) {
+    return NextResponse.json({ error: error.message, code: 'unavailable' }, { status: 503 });
+  }
+  if (error instanceof ReplicaSyncError) {
+    // A transport failure here means the BACKEND is unreachable, which for a sync
+    // is an expected outcome rather than a server fault - 503 with the class, so
+    // the renderer can retry rather than surface an error.
+    const status = error.cls === 'Auth' ? 401 : error.cls === 'RateLimit' ? 429 : 503;
+    return NextResponse.json({ error: error.message, code: error.cls }, { status });
+  }
+  logger.error(`offline-replica: ${context} failed`, {
+    error: error instanceof Error ? error.message : String(error),
+  });
+  return NextResponse.json({ error: `${context} failed` }, { status: 500 });
+}
+
+export const NO_STORE = { 'Cache-Control': 'no-store' } as const;
diff --git a/lib/offline-replica/schema.ts b/lib/offline-replica/schema.ts
new file mode 100644
index 00000000..755e0add
--- /dev/null
+++ b/lib/offline-replica/schema.ts
@@ -0,0 +1,178 @@
+// The offline mail replica's schema.
+//
+// WHY IT LIVES IN THE SAME ENCRYPTED FILE AS THE SEARCH INDEX
+// (`lib/mail-index/paths.ts`'s `indexDbPath`), on its own connection:
+//
+//  * One encryption boundary, one key, one keychain entry, one purge. A second
+//    keystore would double the number of places a key can be mishandled for no
+//    gain - and `electron/key-service.ts` + the fd-3 channel already work.
+//  * `sync_state` (cursors, coverage, flags) sits in the SAME FILE as the
+//    records it describes. That is load-bearing, not tidiness: deleting the file
+//    removes cursors and records together, so a cursor can never survive a wipe
+//    and then be advanced over changes that will never be re-delivered. A
+//    cursor in a sidecar JSON file is exactly the class of bug the mobile
+//    client's S1 finding is about.
+//  * The tables are DISJOINT from the index's (`doc`, `doc_fts`), so the two
+//    subsystems never contend for a row - only, briefly, for SQLite's write
+//    lock, which `PRAGMA busy_timeout` resolves. Both open with WAL.
+//
+// The one coupling to accept: `lib/mail-index/store.ts` drops `meta` on an index
+// schema bump, which takes `replica_schema_version` with it. The replica reads
+// that as "version missing" and purges + re-bootstraps - correct, just wasteful,
+// and only on an index schema change. What must NOT happen is records surviving
+// while the version row vanishes, which is why the purge below is all-or-nothing
+// and includes `replica_sync_state`.
+//
+// Deliberately NO foreign keys from `replica_email_mailbox.mailbox_id` ->
+// `replica_mailbox`, and no cascade from `replica_envelope`. The two change
+// streams are not transactionally coupled, so a membership row referencing a
+// not-yet-fetched or already-destroyed mailbox is a NORMAL transient state; an
+// FK would turn correct behaviour into a constraint violation, and a cascade on
+// mailbox deletion would delete mail, violating the deletion-provenance rule
+// (only the Email stream may delete an email).
+
+/** Bumped on any incompatible change. Mismatch = purge + re-bootstrap, never migrate. */
+export const REPLICA_SCHEMA_VERSION = 1;
+
+export const REPLICA_VERSION_KEY = 'replica_schema_version';
+
+export const REPLICA_DDL = `
+CREATE TABLE IF NOT EXISTS replica_mailbox (
+  jmap_account_id TEXT NOT NULL,
+  id              TEXT NOT NULL,
+  name            TEXT NOT NULL DEFAULT '',
+  parent_id       TEXT,
+  role            TEXT,
+  sort_order      INTEGER,
+  total_emails    INTEGER,
+  unread_emails   INTEGER,
+  total_threads   INTEGER,
+  unread_threads  INTEGER,
+  my_rights_json  TEXT,
+  is_subscribed   INTEGER,
+  PRIMARY KEY (jmap_account_id, id)
+);
+
+-- The envelope tier: everything EMAIL_LIST_PROPERTIES carries, so an offline
+-- message list renders exactly as an online one does.
+CREATE TABLE IF NOT EXISTS replica_envelope (
+  jmap_account_id TEXT NOT NULL,
+  id              TEXT NOT NULL,
+  thread_id       TEXT,
+  received_at     TEXT NOT NULL,
+  size            INTEGER,
+  subject         TEXT,
+  preview         TEXT,
+  from_json       TEXT,
+  to_json         TEXT,
+  cc_json         TEXT,
+  blob_id         TEXT,
+  has_attachment  INTEGER NOT NULL DEFAULT 0,
+  keywords_json   TEXT NOT NULL DEFAULT '{}',
+  -- Owned by the BODY tier. Excluded from the envelope upsert's DO UPDATE SET,
+  -- or an idempotent page replay would look like "body missing" to the backfill
+  -- job and re-download every body in the page.
+  has_body        INTEGER NOT NULL DEFAULT 0,
+  body_bytes      INTEGER NOT NULL DEFAULT 0,
+  cached_at       INTEGER NOT NULL,
+  PRIMARY KEY (jmap_account_id, id)
+);
+CREATE INDEX IF NOT EXISTS replica_envelope_received
+  ON replica_envelope(jmap_account_id, received_at DESC);
+-- The body-backfill driver: envelopes inside the body window with no body yet.
+CREATE INDEX IF NOT EXISTS replica_envelope_nobody
+  ON replica_envelope(jmap_account_id, has_body, received_at DESC);
+
+-- Membership is its own table: an email is in many mailboxes, and listing by
+-- folder must be an index seek rather than a scan of every cached row.
+CREATE TABLE IF NOT EXISTS replica_email_mailbox (
+  jmap_account_id TEXT NOT NULL,
+  email_id        TEXT NOT NULL,
+  mailbox_id      TEXT NOT NULL,
+  PRIMARY KEY (jmap_account_id, email_id, mailbox_id)
+);
+CREATE INDEX IF NOT EXISTS replica_email_mailbox_by_mailbox
+  ON replica_email_mailbox(jmap_account_id, mailbox_id);
+
+-- The body tier. "received_at" is duplicated here on purpose so cap eviction is
+-- a single-table ordered scan that cannot be blinded by a missing join.
+CREATE TABLE IF NOT EXISTS replica_body (
+  jmap_account_id TEXT NOT NULL,
+  email_id        TEXT NOT NULL,
+  received_at     TEXT NOT NULL,
+  json            TEXT NOT NULL,
+  bytes           INTEGER NOT NULL,
+  PRIMARY KEY (jmap_account_id, email_id)
+);
+CREATE INDEX IF NOT EXISTS replica_body_received
+  ON replica_body(jmap_account_id, received_at ASC);
+
+-- "gave_up" is what makes a body-tier TERMINAL STATE durable, and it is the fix
+-- for the worst bug found on the mobile client. Deleting the queue row on
+-- give-up was not enough: the backfill job's driver is "envelope with no body",
+-- a predicate that CANNOT distinguish "not fetched yet" from "deliberately not
+-- kept". So the next pass re-inserted a fresh attempts=0 row and a
+-- permanently-failing body was retried five times per cycle, forever. Same
+-- shape for a "notFound" body, and worst of all for a body shed by the size cap:
+-- shed -> still inside the body window -> re-enqueued -> re-downloaded -> shed
+-- again. Unbounded data use with no termination. Keeping the row with a flag is
+-- what closes all three.
+CREATE TABLE IF NOT EXISTS replica_body_queue (
+  jmap_account_id TEXT NOT NULL,
+  email_id        TEXT NOT NULL,
+  received_at     TEXT NOT NULL,
+  attempts        INTEGER NOT NULL DEFAULT 0,
+  next_attempt_at INTEGER,
+  last_error      TEXT,
+  gave_up         INTEGER NOT NULL DEFAULT 0,
+  gave_up_reason  TEXT,
+  PRIMARY KEY (jmap_account_id, email_id)
+);
+CREATE INDEX IF NOT EXISTS replica_body_queue_wanted
+  ON replica_body_queue(jmap_account_id, gave_up, received_at DESC);
+
+-- Cursors, coverage and flags. Row-per-field, NOT one JSON blob: a blob loaded
+-- at cycle start and written at cycle end silently reverts any concurrent write
+-- to a different field. On the mobile client that produced an empty record store
+-- with a live advanced cursor and resyncRequired reset to false - a permanent
+-- silent data gap that is unreachable by design.
+CREATE TABLE IF NOT EXISTS replica_sync_state (
+  k TEXT PRIMARY KEY,
+  v TEXT NOT NULL
+);
+`;
+
+/**
+ * Everything a purge removes. `replica_sync_state` is INCLUDED: a record wipe
+ * that leaves cursors behind is the one state from which no amount of syncing
+ * recovers, because `/changes` structurally cannot re-deliver mail that already
+ * existed when the cursor was captured.
+ */
+export const REPLICA_TABLES: readonly string[] = [
+  'replica_envelope',
+  'replica_email_mailbox',
+  'replica_body',
+  'replica_body_queue',
+  'replica_mailbox',
+  'replica_sync_state',
+];
+
+/** Record tables only - used when a reconcile rebuilds without discarding policy. */
+export const REPLICA_RECORD_TABLES: readonly string[] = [
+  'replica_envelope',
+  'replica_email_mailbox',
+  'replica_body',
+  'replica_body_queue',
+  'replica_mailbox',
+];
+
+export const FLAGS_KEY = 'flags';
+export const POLICY_KEY = 'policy';
+
+export function cursorStateKey(jmapAccountId: string, type: string): string {
+  return `cursor:${jmapAccountId}:${type}`;
+}
+
+export function coverageStateKey(jmapAccountId: string): string {
+  return `coverage:${jmapAccountId}`;
+}
diff --git a/lib/offline-replica/states.ts b/lib/offline-replica/states.ts
new file mode 100644
index 00000000..429b0de2
--- /dev/null
+++ b/lib/offline-replica/states.ts
@@ -0,0 +1,130 @@
+// Cursor provenance: the type-level machinery that makes "never adopt an
+// `Email/get` state as an `Email/changes` cursor" a compile error rather than a
+// code-review convention.
+//
+// This exact bug shipped on the mobile client (its defect D4) and silently
+// corrupted sync: `getEmailChanges` returned `null` for ANY error, so a
+// transient 503 on `Email/changes` caused an `Email/get` state captured in the
+// same cycle to be adopted as the next cursor - fast-forwarding the cursor over
+// every change the client had not seen, with no resync. The cost is invisible:
+// the store looks healthy and is permanently missing mail.
+//
+// Two brands, and an ORDERING rule rather than a source rule. "Only a
+// `Foo/changes.newState` may ever be a cursor" is tempting but FALSE - bootstrap
+// and reconcile legitimately seed from `Foo/get {ids: []}`'s `state`, which
+// RFC 8620 s5.1 explicitly permits. A rule the design itself has to violate is a
+// rule that gets bypassed at the one call site that matters, so the rule is:
+//
+//   A cursor ADVANCES to a ChangesState from the same (jmapAccountId, type).
+//   It may be SEEDED from a SnapshotState only inside an EnumerationCommitment
+//   whose enumeration starts after that snapshot. Nothing else, from anywhere,
+//   ever becomes a cursor.
+
+/** Types we hold a `/changes` cursor for. NOT a list of push types. */
+export type CursorType = 'Email' | 'Mailbox';
+
+export const CURSOR_TYPES: readonly CursorType[] = ['Email', 'Mailbox'];
+
+/** From a `Foo/changes` response's `newState`. The only value the delta path may advance to. */
+export type ChangesState = string & { readonly __brand: 'ChangesState' };
+
+/** From a `Foo/get` response's `state`. A valid cursor ONLY under the ordering rule above. */
+export type SnapshotState = string & { readonly __brand: 'SnapshotState' };
+
+/**
+ * A JMAP body is parsed JSON, so without a runtime check a `null`, a number or
+ * an object could be laundered through a cast into something the engine treats
+ * as a cursor forever. The brand certifies PROVENANCE; this certifies SHAPE.
+ */
+function certifyStateToken(value: unknown, kind: string): string {
+  if (typeof value !== 'string' || value.length === 0) {
+    throw new TypeError(
+      `${kind}: expected a non-empty string state token, got ` +
+        `${value === null ? 'null' : typeof value}`,
+    );
+  }
+  return value;
+}
+
+/**
+ * Mint a `ChangesState`. Callable ONLY from the `Foo/changes` response parser in
+ * `./jmap.ts` - that is the entire point of the brand. There is a test asserting
+ * no other module casts to these types.
+ */
+export function asChangesState(newState: unknown): ChangesState {
+  return certifyStateToken(newState, 'asChangesState') as ChangesState;
+}
+
+/** Mint a `SnapshotState`. Callable ONLY from the `Foo/get` response parser in `./jmap.ts`. */
+export function asSnapshotState(state: unknown): SnapshotState {
+  return certifyStateToken(state, 'asSnapshotState') as SnapshotState;
+}
+
+/**
+ * The tag is a REAL, module-private `Symbol()`, deliberately not exported and
+ * deliberately not `declare const ... : unique symbol`.
+ *
+ * - Unexported means no object literal in any other module can produce this
+ *   type, so `mintEnumerationCommitment` is the only constructor. Declaring the
+ *   interface without a symbol tag would let any module falsify it with a
+ *   literal, making the seed path's teeth strictly weaker than
+ *   `advanceCursor`'s - which is the path that needs them most.
+ * - `declare const x: unique symbol` is TYPE-LEVEL ONLY and emits no runtime
+ *   value, so using it as a computed key throws
+ *   `ReferenceError: x is not defined` the first time the mint runs. That
+ *   mistake is in the superseded design document; it cost the mobile port a
+ *   build failure. A `Symbol()` assigned to a `const` still infers
+ *   `unique symbol`, so unforgeability is identical and no cast is needed.
+ */
+const enumerationCommitmentTag = Symbol('EnumerationCommitment');
+
+/**
+ * A durable promise to enumerate. Holding one is what entitles a caller to seed
+ * a cursor from a snapshot state: the snapshot is only a safe cursor because an
+ * enumeration that starts AFTER it is committed to run.
+ */
+export interface EnumerationCommitment {
+  readonly [enumerationCommitmentTag]: true;
+  readonly jmapAccountId: string;
+  readonly snapshot: SnapshotState;
+  /** The retention floor the enumeration is working toward. */
+  readonly targetFrom: string;
+  /**
+   * The floor PINNED for this enumeration. Equal to `targetFrom` for a
+   * bootstrap; for a reconcile it is the floor captured at step 0, and the sweep
+   * deletes only against THIS value, never a `targetFrom` that moved while the
+   * reconcile was running.
+   *
+   * Without the pin: widening retention mid-reconcile (very plausible - the
+   * reconcile banner is exactly what prompts someone to go change the setting)
+   * makes the sweep delete against the new wide window while the enumeration
+   * only covered the old narrow one. Everything in the gap is deleted
+   * permanently, because `coveredFrom` is then set to the wider floor and
+   * `/changes` cannot re-deliver pre-existing mail.
+   */
+  readonly sweepFloor: string;
+  readonly kind: 'bootstrap' | 'reconcile';
+}
+
+export function mintEnumerationCommitment(args: {
+  jmapAccountId: string;
+  snapshot: SnapshotState;
+  targetFrom: string;
+  sweepFloor: string;
+  kind: 'bootstrap' | 'reconcile';
+}): EnumerationCommitment {
+  return {
+    [enumerationCommitmentTag]: true,
+    jmapAccountId: args.jmapAccountId,
+    snapshot: args.snapshot,
+    targetFrom: args.targetFrom,
+    sweepFloor: args.sweepFloor,
+    kind: args.kind,
+  };
+}
+
+export function coveragePhaseForCommitment(
+  commitment: EnumerationCommitment,
+): 'scanning' | 'reconciling' {
+  return commitment.kind === 'bootstrap' ? 'scanning' : 'reconciling';
+}
diff --git a/lib/offline-replica/store.ts b/lib/offline-replica/store.ts
new file mode 100644
index 00000000..f95b71d0
--- /dev/null
+++ b/lib/offline-replica/store.ts
@@ -0,0 +1,997 @@
+// The replica store: the encrypted SQLite file, and every read/write against it.
+//
+// SYNCHRONOUS ON PURPOSE. `@signalapp/sqlcipher` is a synchronous binding, so
+// `transaction()` here takes a synchronous callback and no `await` can ever
+// appear inside a `BEGIN ... COMMIT`. That removes an entire hazard class by
+// construction: no network call, no timer and no other request can interleave
+// with a half-applied transaction. Every JMAP fetch happens OUTSIDE a
+// transaction and the results are applied inside one.
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { loadSqlcipher, type SqlcipherDatabase } from '@/lib/mail-index/binding';
+import { dbSiblings, indexDbPath } from '@/lib/mail-index/paths';
+import {
+  coverageStateKey, cursorStateKey, FLAGS_KEY, POLICY_KEY, REPLICA_DDL,
+  REPLICA_RECORD_TABLES, REPLICA_SCHEMA_VERSION, REPLICA_TABLES, REPLICA_VERSION_KEY,
+} from './schema';
+import {
+  coveragePhaseForCommitment, type ChangesState, type CursorType, type EnumerationCommitment,
+} from './states';
+import {
+  defaultFlags, type BodyGiveUpReason, type BodyQueueEntry, type CoverageState,
+  type EnvelopeRow, type FlagsPatch, type MailboxRow, type ReplicaFlags, type SyncCursor,
+} from './types';
+
+export class ReplicaUnavailableError extends Error {
+  constructor(message: string) {
+    super(message);
+    this.name = 'ReplicaUnavailableError';
+  }
+}
+
+export interface CursorKey {
+  jmapAccountId: string;
+  type: CursorType;
+}
+
+/** Retention policy, persisted server-side inside the encrypted store. */
+export interface RetentionPolicy {
+  envelopeDays: number;
+  bodyDays: number;
+  maxBodyMB: number;
+}
+
+export const DEFAULT_POLICY: RetentionPolicy = {
+  // Envelopes are ~1 KB, so a wide window costs kilobytes per message and means
+  // a message never falls out of the offline LIST because of a body size cap.
+  envelopeDays: 180,
+  bodyDays: 30,
+  maxBodyMB: 250,
+};
+
+export const POLICY_LIMITS = {
+  envelopeDays: { min: 7, max: 3650 },
+  bodyDays: { min: 1, max: 3650 },
+  maxBodyMB: { min: 16, max: 20_000 },
+} as const;
+
+export function clampPolicy(raw: Partial<RetentionPolicy> | null | undefined): RetentionPolicy {
+  const pick = (
+    value: unknown,
+    fallback: number,
+    { min, max }: { min: number; max: number },
+  ): number => {
+    const n = typeof value === 'number' && Number.isFinite(value) ? Math.round(value) : fallback;
+    return Math.min(Math.max(n, min), max);
+  };
+  const envelopeDays = pick(raw?.envelopeDays, DEFAULT_POLICY.envelopeDays, POLICY_LIMITS.envelopeDays);
+  const bodyDays = pick(raw?.bodyDays, DEFAULT_POLICY.bodyDays, POLICY_LIMITS.bodyDays);
+  return {
+    envelopeDays,
+    // The body window can never be wider than the envelope window: a body with
+    // no envelope is an orphan by construction.
+    bodyDays: Math.min(bodyDays, envelopeDays),
+    maxBodyMB: pick(raw?.maxBodyMB, DEFAULT_POLICY.maxBodyMB, POLICY_LIMITS.maxBodyMB),
+  };
+}
+
+/**
+ * `PRAGMA cipher_version` must return a non-empty STRING.
+ *
+ * Checking the row COUNT instead passes vacuously: a non-SQLCipher binding
+ * returns ZERO ROWS for this pragma, and `PRAGMA key = ...` is silently accepted
+ * and does nothing on plain SQLite - no error, a working database, and the mail
+ * sitting on disk in cleartext.
+ */
+function assertEncrypted(db: SqlcipherDatabase, dbPath: string): void {
+  const rows = db.pragma('cipher_version');
+  const value =
+    Array.isArray(rows) && rows.length > 0 && rows[0] && typeof rows[0] === 'object'
+      ? (rows[0] as Record<string, unknown>).cipher_version
+      : undefined;
+  if (typeof value !== 'string' || value.trim().length === 0) {
+    db.close();
+    throw new ReplicaUnavailableError(
+      `Refusing to use ${path.basename(dbPath)}: the SQLite binding reports no SQLCipher ` +
+        `support (PRAGMA cipher_version returned ${JSON.stringify(rows)}), so the offline ` +
+        `replica would be written in cleartext.`,
+    );
+  }
+}
+
+function num(v: unknown): number | null {
+  return typeof v === 'number' && Number.isFinite(v) ? v : null;
+}
+
+function str(v: unknown): string | null {
+  return typeof v === 'string' ? v : null;
+}
+
+export interface OpenReplicaOptions {
+  storeDir: string;
+  accountId: string;
+  /** Raw 32-byte key, from the main process's key service. */
+  key: Buffer;
+}
+
+export class ReplicaStore {
+  private constructor(
+    private readonly db: SqlcipherDatabase,
+    readonly dbPath: string,
+  ) {}
+
+  static open({ storeDir, accountId, key }: OpenReplicaOptions): ReplicaStore {
+    const Database = loadSqlcipher();
+    if (!Database) {
+      throw new ReplicaUnavailableError(
+        '@signalapp/sqlcipher is not installed for this platform (it is an optional dependency).',
+      );
+    }
+    if (key.length !== 32) {
+      throw new ReplicaUnavailableError(`Replica key must be 32 bytes, got ${key.length}.`);
+    }
+
+    // The SAME file as the search index. See schema.ts for why.
+    const dbPath = indexDbPath(storeDir, accountId);
+    fs.mkdirSync(path.dirname(dbPath), { recursive: true, mode: 0o700 });
+
+    const connect = (): SqlcipherDatabase => {
+      const db = new Database(dbPath);
+      // The key pragma must be the FIRST statement on the connection. Hex form
+      // means SQLCipher uses these 32 bytes as the raw key with no KDF.
+      db.pragma(`key = "x'${key.toString('hex')}'"`);
+      assertEncrypted(db, dbPath);
+      return db;
+    };
+
+    let db = connect();
+    let version: number | null;
+    try {
+      db.pragma('journal_mode = WAL');
+      db.pragma('synchronous = NORMAL');
+      // The index and the replica are two connections to one file. WAL lets a
+      // writer and readers coexist, but two WRITERS get SQLITE_BUSY immediately
+      // without this - and both are driven by the same renderer push handler, so
+      // they genuinely do overlap.
+      db.pragma('busy_timeout = 8000');
+      version = readVersion(db);
+    } catch {
+      // A wrong key surfaces here, not at open: SQLCipher reads the header
+      // lazily. The replica is derived data, so there is nothing to recover and
+      // never anything to prompt the user for.
+      db.close();
+      for (const f of dbSiblings(dbPath)) {
+        try { fs.rmSync(f, { force: true }); } catch { /* best effort */ }
+      }
+      db = connect();
+      db.pragma('journal_mode = WAL');
+      db.pragma('synchronous = NORMAL');
+      db.pragma('busy_timeout = 8000');
+      version = null;
+    }
+
+    if (version !== null && version !== REPLICA_SCHEMA_VERSION) version = null;
+
+    if (version === null) {
+      // ALL-OR-NOTHING. Records must never survive while the version row is
+      // gone: a cursor that outlives its records is the one state no amount of
+      // syncing repairs, because `/changes` cannot re-deliver mail that already
+      // existed when the cursor was captured.
+      db.exec('BEGIN');
+      try {
+        for (const table of REPLICA_TABLES) db.exec(`DROP TABLE IF EXISTS ${table}`);
+        db.exec(REPLICA_DDL);
+        db.exec('CREATE TABLE IF NOT EXISTS meta (k TEXT PRIMARY KEY, v TEXT NOT NULL)');
+        db.prepare('INSERT OR REPLACE INTO meta (k, v) VALUES (?, ?)').run([
+          REPLICA_VERSION_KEY,
+          String(REPLICA_SCHEMA_VERSION),
+        ]);
+        db.exec('COMMIT');
+      } catch (error) {
+        db.exec('ROLLBACK');
+        db.close();
+        throw error;
+      }
+    } else {
+      // The tables exist per the version row, but `CREATE TABLE IF NOT EXISTS`
+      // is cheap and covers a partially-created file from an interrupted open.
+      db.exec(REPLICA_DDL);
+    }
+
+    return new ReplicaStore(db, dbPath);
+  }
+
+  close(): void {
+    try { this.db.close(); } catch { /* already closed */ }
+  }
+
+  /**
+   * One SQLite transaction. The callback is SYNCHRONOUS, so nothing can
+   * interleave and no `await` can sit inside `BEGIN ... COMMIT`.
+   */
+  transaction<T>(fn: () => T): T {
+    this.db.exec('BEGIN');
+    try {
+      const out = fn();
+      this.db.exec('COMMIT');
+      return out;
+    } catch (error) {
+      try { this.db.exec('ROLLBACK'); } catch { /* the commit may have failed */ }
+      throw error;
+    }
+  }
+
+  // ── raw sync_state access ────────────────────────────────────────────────
+
+  private readState<T>(k: string): T | null {
+    const row = this.db.prepare('SELECT v FROM replica_sync_state WHERE k = ?').get([k]);
+    if (!row || typeof row.v !== 'string') return null;
+    try {
+      return JSON.parse(row.v) as T;
+    } catch {
+      // A corrupt state blob is a resync signal, not something to guess at.
+      return null;
+    }
+  }
+
+  private writeState(k: string, value: unknown): void {
+    this.db
+      .prepare('INSERT INTO replica_sync_state (k, v) VALUES (?, ?) ON CONFLICT(k) DO UPDATE SET v = excluded.v')
+      .run([k, JSON.stringify(value)]);
+  }
+
+  // ── policy ───────────────────────────────────────────────────────────────
+
+  getPolicy(): RetentionPolicy {
+    return clampPolicy(this.readState<Partial<RetentionPolicy>>(POLICY_KEY));
+  }
+
+  setPolicy(policy: RetentionPolicy): void {
+    this.writeState(POLICY_KEY, clampPolicy(policy));
+  }
+
+  // ── flags ────────────────────────────────────────────────────────────────
+
+  getFlags(now: number): ReplicaFlags {
+    return this.readState<ReplicaFlags>(FLAGS_KEY) ?? defaultFlags(now);
+  }
+
+  patchFlags(now: number, patch: FlagsPatch): void {
+    const current = this.getFlags(now);
+    this.writeState(FLAGS_KEY, { ...current, ...patch });
+  }
+
+  // ── cursors ──────────────────────────────────────────────────────────────
+
+  getCursor(key: CursorKey): SyncCursor | null {
+    return this.readState<SyncCursor>(cursorStateKey(key.jmapAccountId, key.type));
+  }
+
+  /**
+   * The delta path's ONLY cursor write. The signature is what makes the mobile
+   * client's D4 a compile error here: a `SnapshotState` cannot be passed.
+   *
+   * Throws when the cursor does not exist. A cursor is born from `seedCursor`
+   * and nowhere else; creating one here would be a silent cursor-from-nowhere,
+   * which is the exact class of bug the branded types exist to prevent.
+   */
+  advanceCursor(key: CursorKey, next: ChangesState): void {
+    const k = cursorStateKey(key.jmapAccountId, key.type);
+    const current = this.readState<SyncCursor>(k);
+    if (!current) {
+      throw new Error(
+        `advanceCursor: no cursor for ${key.type}/${key.jmapAccountId}; seed it first`,
+      );
+    }
+    this.writeState(k, { ...current, state: next, updatedAt: Date.now() } satisfies SyncCursor);
+  }
+
+  /**
+   * Bootstrap / reconcile only. Writes the snapshot state AND the `CoverageState`
+   * it justifies in the SAME transaction, so a seed is never durable without the
+   * durable commitment to enumerate that justifies it.
+   *
+   * Call inside `transaction()`.
+   */
+  seedCursor(key: CursorKey, commitment: EnumerationCommitment, now: number): void {
+    if (commitment.jmapAccountId !== key.jmapAccountId) {
+      throw new Error('seedCursor: commitment is for a different JMAP account');
+    }
+    const seeded: SyncCursor = {
+      type: key.type,
+      jmapAccountId: key.jmapAccountId,
+      state: commitment.snapshot,
+      drainPending: false,
+      consecutiveFailures: 0,
+      maxChangesRung: 0,
+      updatedAt: now,
+    };
+    this.writeState(cursorStateKey(key.jmapAccountId, key.type), seeded);
+
+    const existing = this.getCoverage(key.jmapAccountId);
+    const next: CoverageState = {
+      jmapAccountId: key.jmapAccountId,
+      // Records stay readable during a reconcile, so what was already covered
+      // stays claimed until the reconcile finishes and sets the pinned floor.
+      coveredFrom: existing?.coveredFrom ?? null,
+      scanCursor: null,
+      targetFrom: commitment.targetFrom,
+      sweepFloor: commitment.sweepFloor,
+      deferredTargetFrom: undefined,
+      gapMarkers: existing?.gapMarkers,
+      phase: coveragePhaseForCommitment(commitment),
+      seen: 0,
+      consecutiveFailures: 0,
+      updatedAt: now,
+    };
+    this.writeState(coverageStateKey(key.jmapAccountId), next);
+  }
+
+  /** Field-level patch. `state` is deliberately NOT patchable - see advance/seed. */
+  patchCursor(
+    key: CursorKey,
+    patch: Partial<Omit<SyncCursor, 'type' | 'jmapAccountId' | 'state'>>,
+  ): void {
+    const k = cursorStateKey(key.jmapAccountId, key.type);
+    const current = this.readState<SyncCursor>(k);
+    if (!current) return;
+    this.writeState(k, { ...current, ...patch, updatedAt: Date.now() });
+  }
+
+  // ── coverage ─────────────────────────────────────────────────────────────
+
+  getCoverage(jmapAccountId: string): CoverageState | null {
+    return this.readState<CoverageState>(coverageStateKey(jmapAccountId));
+  }
+
+  patchCoverage(jmapAccountId: string, patch: Partial<CoverageState>): void {
+    const k = coverageStateKey(jmapAccountId);
+    const current = this.readState<CoverageState>(k);
+    if (!current) return;
+    this.writeState(k, { ...current, ...patch, updatedAt: Date.now() });
+  }
+
+  // ── mailboxes ────────────────────────────────────────────────────────────
+
+  upsertMailboxes(rows: readonly MailboxRow[]): number {
+    if (rows.length === 0) return 0;
+    const stmt = this.db.prepare(`
+      INSERT INTO replica_mailbox (jmap_account_id, id, name, parent_id, role, sort_order,
+        total_emails, unread_emails, total_threads, unread_threads, my_rights_json, is_subscribed)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+      ON CONFLICT(jmap_account_id, id) DO UPDATE SET
+        name = excluded.name, parent_id = excluded.parent_id, role = excluded.role,
+        sort_order = excluded.sort_order, total_emails = excluded.total_emails,
+        unread_emails = excluded.unread_emails, total_threads = excluded.total_threads,
+        unread_threads = excluded.unread_threads, my_rights_json = excluded.my_rights_json,
+        is_subscribed = excluded.is_subscribed
+    `);
+    for (const r of rows) {
+      stmt.run([
+        r.jmapAccountId, r.id, r.name, r.parentId, r.role, r.sortOrder,
+        r.totalEmails, r.unreadEmails, r.totalThreads, r.unreadThreads,
+        r.myRightsJson, r.isSubscribed ? 1 : 0,
+      ]);
+    }
+    return rows.length;
+  }
+
+  /**
+   * Patches ONLY the four count columns.
+   *
+   * `Mailbox/changes` reports `updatedProperties` as an upper bound of what may
+   * have changed (RFC 8621 s2.2), and counts move on every delivery and every
+   * read. On a busy account this is the difference between patching four
+   * integers and re-fetching every folder object.
+   */
+  patchMailboxCounts(
+    jmapAccountId: string,
+    id: string,
+    counts: {
+      totalEmails?: number | null; unreadEmails?: number | null;
+      totalThreads?: number | null; unreadThreads?: number | null;
+    },
+  ): void {
+    const sets: string[] = [];
+    const params: unknown[] = [];
+    for (const [column, value] of [
+      ['total_emails', counts.totalEmails], ['unread_emails', counts.unreadEmails],
+      ['total_threads', counts.totalThreads], ['unread_threads', counts.unreadThreads],
+    ] as const) {
+      if (value !== undefined) { sets.push(`${column} = ?`); params.push(value); }
+    }
+    if (sets.length === 0) return;
+    this.db
+      .prepare(`UPDATE replica_mailbox SET ${sets.join(', ')} WHERE jmap_account_id = ? AND id = ?`)
+      .run([...params, jmapAccountId, id]);
+  }
+
+  /**
+   * Deletes the mailbox row ONLY. Never touches email records.
+   *
+   * Deletion provenance: if the server destroyed the messages too,
+   * `Email/changes` reports them `destroyed`; if it moved them, their
+   * `mailboxIds` update arrives as `updated`. Truth arrives on the Email stream
+   * either way. Inferring deletion from a mailbox disappearing is how a client
+   * loses mail the server still has.
+   */
+  deleteMailboxes(jmapAccountId: string, ids: readonly string[]): number {
+    if (ids.length === 0) return 0;
+    const stmt = this.db.prepare('DELETE FROM replica_mailbox WHERE jmap_account_id = ? AND id = ?');
+    let n = 0;
+    for (const id of ids) n += stmt.run([jmapAccountId, id]).changes;
+    return n;
+  }
+
+  listMailboxes(jmapAccountId: string): MailboxRow[] {
+    return this.db
+      .prepare('SELECT * FROM replica_mailbox WHERE jmap_account_id = ? ORDER BY sort_order ASC, name ASC')
+      .all([jmapAccountId])
+      .map((r) => ({
+        jmapAccountId: String(r.jmap_account_id),
+        id: String(r.id),
+        name: String(r.name ?? ''),
+        parentId: str(r.parent_id),
+        role: str(r.role),
+        sortOrder: num(r.sort_order),
+        totalEmails: num(r.total_emails),
+        unreadEmails: num(r.unread_emails),
+        totalThreads: num(r.total_threads),
+        unreadThreads: num(r.unread_threads),
+        myRightsJson: str(r.my_rights_json),
+        isSubscribed: r.is_subscribed !== 0,
+      }));
+  }
+
+  // ── envelopes ────────────────────────────────────────────────────────────
+
+  /**
+   * Upserts envelopes and replaces their membership rows.
+   *
+   * `has_body` / `body_bytes` are DELIBERATELY absent from `DO UPDATE SET`: they
+   * belong to the body tier, and resetting them on an idempotent page replay
+   * would look like "body missing" to the backfill job and re-download every
+   * body in the page.
+   *
+   * `cachedAt` is a parameter rather than `Date.now()` because a reconcile must
+   * stamp with its PINNED value - see `CoverageState.reconcileStampedAt`.
+   */
+  upsertEnvelopes(rows: readonly EnvelopeRow[], cachedAt: number): number {
+    if (rows.length === 0) return 0;
+    const upsert = this.db.prepare(`
+      INSERT INTO replica_envelope (jmap_account_id, id, thread_id, received_at, size, subject,
+        preview, from_json, to_json, cc_json, blob_id, has_attachment, keywords_json,
+        has_body, body_bytes, cached_at)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 0, 0, ?)
+      ON CONFLICT(jmap_account_id, id) DO UPDATE SET
+        thread_id = excluded.thread_id, received_at = excluded.received_at,
+        size = excluded.size, subject = excluded.subject, preview = excluded.preview,
+        from_json = excluded.from_json, to_json = excluded.to_json, cc_json = excluded.cc_json,
+        blob_id = excluded.blob_id, has_attachment = excluded.has_attachment,
+        keywords_json = excluded.keywords_json, cached_at = excluded.cached_at
+    `);
+    const clearMembership = this.db.prepare(
+      'DELETE FROM replica_email_mailbox WHERE jmap_account_id = ? AND email_id = ?',
+    );
+    const addMembership = this.db.prepare(
+      'INSERT OR IGNORE INTO replica_email_mailbox (jmap_account_id, email_id, mailbox_id) VALUES (?, ?, ?)',
+    );
+    for (const r of rows) {
+      upsert.run([
+        r.jmapAccountId, r.id, r.threadId, r.receivedAt, r.size, r.subject, r.preview,
+        r.fromJson, r.toJson, r.ccJson, r.blobId, r.hasAttachment ? 1 : 0, r.keywordsJson,
+        cachedAt,
+      ]);
+      clearMembership.run([r.jmapAccountId, r.id]);
+      for (const mailboxId of r.mailboxIds) addMembership.run([r.jmapAccountId, r.id, mailboxId]);
+    }
+    return rows.length;
+  }
+
+  /**
+   * Patches the only two MUTABLE Email properties (RFC 8621 s4.1): `keywords`
+   * and `mailboxIds`. Everything else - body, attachments, headers, receivedAt,
+   * size, threadId, preview, subject, addresses - is immutable for the lifetime
+   * of the id, which is why an `updated` id never needs a body re-fetch.
+   *
+   * No-ops for an id we do not hold, and only touches membership when the
+   * envelope row actually existed, or we would leave membership rows for a
+   * record we do not have.
+   */
+  patchEnvelopeMutable(
+    jmapAccountId: string,
+    id: string,
+    patch: { keywordsJson: string; mailboxIds: string[] },
+  ): boolean {
+    const res = this.db
+      .prepare('UPDATE replica_envelope SET keywords_json = ? WHERE jmap_account_id = ? AND id = ?')
+      .run([patch.keywordsJson, jmapAccountId, id]);
+    if (res.changes === 0) return false;
+    this.db
+      .prepare('DELETE FROM replica_email_mailbox WHERE jmap_account_id = ? AND email_id = ?')
+      .run([jmapAccountId, id]);
+    const add = this.db.prepare(
+      'INSERT OR IGNORE INTO replica_email_mailbox (jmap_account_id, email_id, mailbox_id) VALUES (?, ?, ?)',
+    );
+    for (const mailboxId of patch.mailboxIds) add.run([jmapAccountId, id, mailboxId]);
+    return true;
+  }
+
+  /** Bulk presence test, so the delta path can filter `updated` ids BEFORE fetching. */
+  whichEnvelopesExist(jmapAccountId: string, ids: readonly string[]): Set<string> {
+    if (ids.length === 0) return new Set();
+    const out = new Set<string>();
+    const stmt = this.db.prepare(
+      'SELECT id FROM replica_envelope WHERE jmap_account_id = ? AND id = ?',
+    );
+    for (const id of ids) {
+      if (stmt.get([jmapAccountId, id])) out.add(id);
+    }
+    return out;
+  }
+
+  /** Deletes an email everywhere: envelope, body, membership and any queue row. */
+  deleteEmails(jmapAccountId: string, ids: readonly string[]): number {
+    if (ids.length === 0) return 0;
+    const statements = [
+      this.db.prepare('DELETE FROM replica_body WHERE jmap_account_id = ? AND email_id = ?'),
+      this.db.prepare('DELETE FROM replica_body_queue WHERE jmap_account_id = ? AND email_id = ?'),
+      this.db.prepare('DELETE FROM replica_email_mailbox WHERE jmap_account_id = ? AND email_id = ?'),
+    ];
+    const deleteEnvelope = this.db.prepare(
+      'DELETE FROM replica_envelope WHERE jmap_account_id = ? AND id = ?',
+    );
+    let n = 0;
+    for (const id of ids) {
+      for (const s of statements) s.run([jmapAccountId, id]);
+      n += deleteEnvelope.run([jmapAccountId, id]).changes;
+    }
+    return n;
+  }
+
+  /** Retention eviction: everything strictly older than the floor. */
+  evictEnvelopesBelow(jmapAccountId: string, isoFloor: string): number {
+    const ids = this.db
+      .prepare('SELECT id FROM replica_envelope WHERE jmap_account_id = ? AND received_at < ?')
+      .all([jmapAccountId, isoFloor])
+      .map((r) => String(r.id));
+    return this.deleteEmails(jmapAccountId, ids);
+  }
+
+  /**
+   * The reconcile sweep. Two clauses, and it REFUSES to run without a pinned
+   * stamp rather than deleting unverified records.
+   */
+  sweep(jmapAccountId: string, sweepFloor: string, reconcileStampedAt: number | undefined): number {
+    if (reconcileStampedAt === undefined) {
+      throw new Error('sweep: no reconcileStampedAt pinned; refusing to delete unverified records');
+    }
+    const notReSeen = this.db
+      .prepare(`
+        SELECT id FROM replica_envelope
+        WHERE jmap_account_id = ? AND received_at >= ? AND cached_at < ?
+      `)
+      .all([jmapAccountId, sweepFloor, reconcileStampedAt])
+      .map((r) => String(r.id));
+    // Records older than the pinned floor cannot be verified by an enumeration
+    // that only covers the window, so they go rather than being kept on faith.
+    // Normally retention has already evicted them.
+    const unverifiable = this.db
+      .prepare('SELECT id FROM replica_envelope WHERE jmap_account_id = ? AND received_at < ?')
+      .all([jmapAccountId, sweepFloor])
+      .map((r) => String(r.id));
+    return this.deleteEmails(jmapAccountId, [...new Set([...notReSeen, ...unverifiable])]);
+  }
+
+  /**
+   * The reconcile stamp must be derived from the DATA, not the clock:
+   * `max(now, maxCachedAt + 1)`. With a frozen or coarse clock,
+   * `cached_at < stamp` matches nothing and the sweep silently deletes nothing.
+   */
+  maxEnvelopeCachedAt(jmapAccountId: string): number {
+    const row = this.db
+      .prepare('SELECT MAX(cached_at) AS m FROM replica_envelope WHERE jmap_account_id = ?')
+      .get([jmapAccountId]);
+    return num(row?.m) ?? 0;
+  }
+
+  countEnvelopes(jmapAccountId: string): number {
+    const row = this.db
+      .prepare('SELECT COUNT(*) AS n FROM replica_envelope WHERE jmap_account_id = ?')
+      .get([jmapAccountId]);
+    return num(row?.n) ?? 0;
+  }
+
+  /** Envelopes inside the body window with no body yet - the backfill driver. */
+  envelopesWithoutBody(
+    jmapAccountId: string,
+    receivedAfter: string,
+    limit: number,
+  ): Array<{ id: string; receivedAt: string; size: number }> {
+    return this.db
+      .prepare(`
+        SELECT id, received_at, size FROM replica_envelope
+        WHERE jmap_account_id = ? AND has_body = 0 AND received_at >= ?
+        ORDER BY received_at DESC LIMIT ?
+      `)
+      .all([jmapAccountId, receivedAfter, limit])
+      .map((r) => ({
+        id: String(r.id),
+        receivedAt: String(r.received_at),
+        size: num(r.size) ?? 0,
+      }));
+  }
+
+  // ── bodies ───────────────────────────────────────────────────────────────
+
+  /**
+   * Writes a body ONLY if its envelope still exists, and returns whether it did.
+   *
+   * Without the condition, a body fetched moments before its envelope was
+   * destroyed in the same cycle lands as an orphan. This is also exactly why
+   * "run bodies in parallel, it's separate state" is forbidden.
+   */
+  putBodyIfEnvelopeExists(jmapAccountId: string, emailId: string, json: string): boolean {
+    const envelope = this.db
+      .prepare('SELECT received_at FROM replica_envelope WHERE jmap_account_id = ? AND id = ?')
+      .get([jmapAccountId, emailId]);
+    const receivedAt = str(envelope?.received_at);
+    if (receivedAt === null) return false;
+    const bytes = Buffer.byteLength(json, 'utf8');
+    this.db
+      .prepare(`
+        INSERT INTO replica_body (jmap_account_id, email_id, received_at, json, bytes)
+        VALUES (?, ?, ?, ?, ?)
+        ON CONFLICT(jmap_account_id, email_id) DO UPDATE SET
+          received_at = excluded.received_at, json = excluded.json, bytes = excluded.bytes
+      `)
+      .run([jmapAccountId, emailId, receivedAt, json, bytes]);
+    this.db
+      .prepare('UPDATE replica_envelope SET has_body = 1, body_bytes = ? WHERE jmap_account_id = ? AND id = ?')
+      .run([bytes, jmapAccountId, emailId]);
+    return true;
+  }
+
+  getBody(jmapAccountId: string, emailId: string): string | null {
+    const row = this.db
+      .prepare('SELECT json FROM replica_body WHERE jmap_account_id = ? AND email_id = ?')
+      .get([jmapAccountId, emailId]);
+    return str(row?.json);
+  }
+
+  deleteBodies(jmapAccountId: string, emailIds: readonly string[]): number {
+    if (emailIds.length === 0) return 0;
+    const del = this.db.prepare(
+      'DELETE FROM replica_body WHERE jmap_account_id = ? AND email_id = ?',
+    );
+    const clearFlag = this.db.prepare(
+      'UPDATE replica_envelope SET has_body = 0, body_bytes = 0 WHERE jmap_account_id = ? AND id = ?',
+    );
+    let n = 0;
+    for (const id of emailIds) {
+      n += del.run([jmapAccountId, id]).changes;
+      clearFlag.run([jmapAccountId, id]);
+    }
+    return n;
+  }
+
+  bodyBytesTotal(jmapAccountId: string): number {
+    const row = this.db
+      .prepare('SELECT COALESCE(SUM(bytes), 0) AS n FROM replica_body WHERE jmap_account_id = ?')
+      .get([jmapAccountId]);
+    return num(row?.n) ?? 0;
+  }
+
+  countBodies(jmapAccountId: string): number {
+    const row = this.db
+      .prepare('SELECT COUNT(*) AS n FROM replica_body WHERE jmap_account_id = ?')
+      .get([jmapAccountId]);
+    return num(row?.n) ?? 0;
+  }
+
+  /** Oldest bodies first - the cap-eviction order. Envelopes always survive. */
+  oldestBodies(jmapAccountId: string, limit: number): Array<{ emailId: string; bytes: number }> {
+    return this.db
+      .prepare(`
+        SELECT email_id, bytes FROM replica_body WHERE jmap_account_id = ?
+        ORDER BY received_at ASC, email_id ASC LIMIT ?
+      `)
+      .all([jmapAccountId, limit])
+      .map((r) => ({ emailId: String(r.email_id), bytes: num(r.bytes) ?? 0 }));
+  }
+
+  /** Bodies below the body-retention floor. */
+  bodiesBelow(jmapAccountId: string, isoFloor: string, limit: number): string[] {
+    return this.db
+      .prepare(`
+        SELECT email_id FROM replica_body
+        WHERE jmap_account_id = ? AND received_at < ? ORDER BY received_at ASC LIMIT ?
+      `)
+      .all([jmapAccountId, isoFloor, limit])
+      .map((r) => String(r.email_id));
+  }
+
+  /** Bodies whose envelope is gone. Invisible to cap eviction, which walks the body table. */
+  orphanBodies(jmapAccountId: string, limit: number): string[] {
+    return this.db
+      .prepare(`
+        SELECT b.email_id FROM replica_body b
+        LEFT JOIN replica_envelope e
+          ON e.jmap_account_id = b.jmap_account_id AND e.id = b.email_id
+        WHERE b.jmap_account_id = ? AND e.id IS NULL LIMIT ?
+      `)
+      .all([jmapAccountId, limit])
+      .map((r) => String(r.email_id));
+  }
+
+  // ── body queue ───────────────────────────────────────────────────────────
+
+  /**
+   * Insert-or-ignore. NEVER resets `attempts` on an existing row, and never
+   * revives a `gave_up` row.
+   *
+   * Returns the number of rows ACTUALLY INSERTED. The distinction matters: the
+   * caller reports this as progress, and reporting attempted-rather-than-inserted
+   * made the mobile engine believe there was unfinished work on every cycle for
+   * as long as any envelope lacked a body - an endless chain of cycles seconds
+   * apart, changing nothing.
+   */
+  enqueueBodies(entries: readonly BodyQueueEntry[]): number {
+    if (entries.length === 0) return 0;
+    const stmt = this.db.prepare(`
+      INSERT OR IGNORE INTO replica_body_queue
+        (jmap_account_id, email_id, received_at, attempts, next_attempt_at, last_error, gave_up, gave_up_reason)
+      VALUES (?, ?, ?, ?, ?, ?, 0, NULL)
+    `);
+    let inserted = 0;
+    for (const e of entries) {
+      inserted += stmt.run([
+        e.jmapAccountId, e.emailId, e.receivedAt, e.attempts, e.nextAttemptAt ?? null,
+        e.lastError ?? null,
+      ]).changes;
+    }
+    return inserted;
+  }
+
+  /** Rows still WANTED: not given up, and past any backoff. Newest first. */
+  takeBodyQueue(jmapAccountId: string, limit: number, now: number): BodyQueueEntry[] {
+    return this.db
+      .prepare(`
+        SELECT * FROM replica_body_queue
+        WHERE jmap_account_id = ? AND gave_up = 0
+          AND (next_attempt_at IS NULL OR next_attempt_at <= ?)
+        ORDER BY received_at DESC LIMIT ?
+      `)
+      .all([jmapAccountId, now, limit])
+      .map((r) => ({
+        emailId: String(r.email_id),
+        jmapAccountId: String(r.jmap_account_id),
+        receivedAt: String(r.received_at),
+        attempts: num(r.attempts) ?? 0,
+        lastError: str(r.last_error) ?? undefined,
+        nextAttemptAt: num(r.next_attempt_at) ?? undefined,
+        gaveUp: r.gave_up !== 0,
+        gaveUpReason: (str(r.gave_up_reason) as BodyGiveUpReason | null) ?? undefined,
+      }));
+  }
+
+  dequeueBodies(jmapAccountId: string, emailIds: readonly string[]): number {
+    if (emailIds.length === 0) return 0;
+    const stmt = this.db.prepare(
+      'DELETE FROM replica_body_queue WHERE jmap_account_id = ? AND email_id = ?',
+    );
+    let n = 0;
+    for (const id of emailIds) n += stmt.run([jmapAccountId, id]).changes;
+    return n;
+  }
+
+  bumpBodyAttempt(
+    jmapAccountId: string,
+    emailId: string,
+    nextAttemptAt: number,
+    lastError: string,
+  ): void {
+    this.db
+      .prepare(`
+        UPDATE replica_body_queue SET attempts = attempts + 1, next_attempt_at = ?, last_error = ?
+        WHERE jmap_account_id = ? AND email_id = ?
+      `)
+      .run([nextAttemptAt, lastError.slice(0, 400), jmapAccountId, emailId]);
+  }
+
+  /** Records a durable terminal state INSTEAD of deleting the row. */
+  markBodyGaveUp(
+    jmapAccountId: string,
+    entries: ReadonlyArray<{ emailId: string; receivedAt: string; reason: BodyGiveUpReason; lastError?: string }>,
+  ): void {
+    if (entries.length === 0) return;
+    // A cap-shed body may have no queue row at all (it was fetched and stored
+    // successfully, then evicted), so this must be an upsert rather than an
+    // update - otherwise the mark is silently dropped and the shed/re-download
+    // loop stays open.
+    const stmt = this.db.prepare(`
+      INSERT INTO replica_body_queue
+        (jmap_account_id, email_id, received_at, attempts, next_attempt_at, last_error, gave_up, gave_up_reason)
+      VALUES (?, ?, ?, 0, NULL, ?, 1, ?)
+      ON CONFLICT(jmap_account_id, email_id) DO UPDATE SET
+        gave_up = 1, gave_up_reason = excluded.gave_up_reason,
+        last_error = excluded.last_error, next_attempt_at = NULL
+    `);
+    for (const e of entries) {
+      stmt.run([jmapAccountId, e.emailId, e.receivedAt, e.lastError?.slice(0, 400) ?? null, e.reason]);
+    }
+  }
+
+  listBodyGiveUps(jmapAccountId: string, limit: number): string[] {
+    return this.db
+      .prepare('SELECT email_id FROM replica_body_queue WHERE jmap_account_id = ? AND gave_up = 1 LIMIT ?')
+      .all([jmapAccountId, limit])
+      .map((r) => String(r.email_id));
+  }
+
+  /**
+   * DELETES give-up rows rather than un-flagging them, so a cleared give-up
+   * looks like "never queued" and the backfill pass re-enqueues it with a clean
+   * attempt count.
+   *
+   * Called unconditionally by a completed reconcile: a give-up recorded during
+   * whatever went wrong must not outlive it, or a transient outage would
+   * permanently deny a body with no path back.
+   */
+  clearBodyGiveUps(jmapAccountId: string, reason?: BodyGiveUpReason): number {
+    if (reason) {
+      return this.db
+        .prepare('DELETE FROM replica_body_queue WHERE jmap_account_id = ? AND gave_up = 1 AND gave_up_reason = ?')
+        .run([jmapAccountId, reason]).changes;
+    }
+    return this.db
+      .prepare('DELETE FROM replica_body_queue WHERE jmap_account_id = ? AND gave_up = 1')
+      .run([jmapAccountId]).changes;
+  }
+
+  countWantedBodies(jmapAccountId: string, now: number): number {
+    const row = this.db
+      .prepare(`
+        SELECT COUNT(*) AS n FROM replica_body_queue
+        WHERE jmap_account_id = ? AND gave_up = 0
+          AND (next_attempt_at IS NULL OR next_attempt_at <= ?)
+      `)
+      .get([jmapAccountId, now]);
+    return num(row?.n) ?? 0;
+  }
+
+  // ── purge ────────────────────────────────────────────────────────────────
+
+  /** Wipes records AND the body queue. Leaves `replica_sync_state` (policy, cursors). */
+  clearRecords(): void {
+    for (const table of REPLICA_RECORD_TABLES) this.db.exec(`DELETE FROM ${table}`);
+  }
+
+  /** Everything, cursors included. The only safe pairing with a record wipe. */
+  purgeAll(): void {
+    for (const table of REPLICA_TABLES) this.db.exec(`DELETE FROM ${table}`);
+  }
+
+  // ── read path ────────────────────────────────────────────────────────────
+
+  /** Envelope page for a mailbox, newest first. `mailboxId === null` = all mail. */
+  listEnvelopes(
+    jmapAccountId: string,
+    mailboxId: string | null,
+    limit: number,
+    offset: number,
+  ): { rows: Array<Record<string, unknown>>; total: number } {
+    if (mailboxId === null) {
+      const total =
+        num(
+          this.db
+            .prepare('SELECT COUNT(*) AS n FROM replica_envelope WHERE jmap_account_id = ?')
+            .get([jmapAccountId])?.n,
+        ) ?? 0;
+      const rows = this.db
+        .prepare(`
+          SELECT * FROM replica_envelope WHERE jmap_account_id = ?
+          ORDER BY received_at DESC, id DESC LIMIT ? OFFSET ?
+        `)
+        .all([jmapAccountId, limit, offset]);
+      return { rows, total };
+    }
+    const total =
+      num(
+        this.db
+          .prepare('SELECT COUNT(*) AS n FROM replica_email_mailbox WHERE jmap_account_id = ? AND mailbox_id = ?')
+          .get([jmapAccountId, mailboxId])?.n,
+      ) ?? 0;
+    const rows = this.db
+      .prepare(`
+        SELECT e.* FROM replica_envelope e
+        JOIN replica_email_mailbox m
+          ON m.jmap_account_id = e.jmap_account_id AND m.email_id = e.id
+        WHERE e.jmap_account_id = ? AND m.mailbox_id = ?
+        ORDER BY e.received_at DESC, e.id DESC LIMIT ? OFFSET ?
+      `)
+      .all([jmapAccountId, mailboxId, limit, offset]);
+    return { rows, total };
+  }
+
+  getEnvelopeRaw(jmapAccountId: string, id: string): Record<string, unknown> | null {
+    const row = this.db
+      .prepare('SELECT * FROM replica_envelope WHERE jmap_account_id = ? AND id = ?')
+      .get([jmapAccountId, id]);
+    return row ?? null;
+  }
+
+  mailboxIdsFor(jmapAccountId: string, emailId: string): string[] {
+    return this.db
+      .prepare('SELECT mailbox_id FROM replica_email_mailbox WHERE jmap_account_id = ? AND email_id = ?')
+      .all([jmapAccountId, emailId])
+      .map((r) => String(r.mailbox_id));
+  }
+
+  /** Size + freshness, for the Settings surface. */
+  stats(jmapAccountId: string): {
+    mailboxes: number;
+    envelopes: number;
+    bodies: number;
+    bodyBytes: number;
+    wantedBodies: number;
+    giveUps: number;
+    newest: string | null;
+    oldest: string | null;
+    fileBytes: number;
+  } {
+    const mailboxes =
+      num(
+        this.db
+          .prepare('SELECT COUNT(*) AS n FROM replica_mailbox WHERE jmap_account_id = ?')
+          .get([jmapAccountId])?.n,
+      ) ?? 0;
+    const range = this.db
+      .prepare('SELECT MIN(received_at) AS lo, MAX(received_at) AS hi FROM replica_envelope WHERE jmap_account_id = ?')
+      .get([jmapAccountId]);
+    let fileBytes = 0;
+    for (const f of dbSiblings(this.dbPath)) {
+      try { fileBytes += fs.statSync(f).size; } catch { /* absent sibling */ }
+    }
+    return {
+      mailboxes,
+      envelopes: this.countEnvelopes(jmapAccountId),
+      bodies: this.countBodies(jmapAccountId),
+      bodyBytes: this.bodyBytesTotal(jmapAccountId),
+      wantedBodies: this.countWantedBodies(jmapAccountId, Date.now()),
+      giveUps:
+        num(
+          this.db
+            .prepare('SELECT COUNT(*) AS n FROM replica_body_queue WHERE jmap_account_id = ? AND gave_up = 1')
+            .get([jmapAccountId])?.n,
+        ) ?? 0,
+      newest: str(range?.hi),
+      oldest: str(range?.lo),
+      fileBytes,
+    };
+  }
+
+  /** Every JMAP account id with rows, so the read path can find them without a session. */
+  knownJmapAccountIds(): string[] {
+    const ids = new Set<string>();
+    for (const table of ['replica_envelope', 'replica_mailbox'] as const) {
+      for (const r of this.db.prepare(`SELECT DISTINCT jmap_account_id FROM ${table}`).all()) {
+        if (typeof r.jmap_account_id === 'string') ids.add(r.jmap_account_id);
+      }
+    }
+    return [...ids];
+  }
+}
+
+function readVersion(db: SqlcipherDatabase): number | null {
+  try {
+    const row = db.prepare('SELECT v FROM meta WHERE k = ?').get([REPLICA_VERSION_KEY]);
+    if (!row || row.v === undefined) return null;
+    const n = Number(row.v);
+    return Number.isFinite(n) ? n : null;
+  } catch {
+    // `meta` doesn't exist yet - a fresh file.
+    return null;
+  }
+}
diff --git a/lib/offline-replica/sync.ts b/lib/offline-replica/sync.ts
new file mode 100644
index 00000000..7d2e7bac
--- /dev/null
+++ b/lib/offline-replica/sync.ts
@@ -0,0 +1,1122 @@
+// One sync cycle. Request-scoped, bounded, single-flighted.
+//
+// PROCESS ARCHITECTURE - the decision that keeps most of the design review's
+// findings out of scope. There is NO persistent background worker. A cycle is
+// ordinary work inside an API route, using the request's own encrypted
+// `jmap_stalwart_ctx` cookie via `lib/stalwart/credentials.ts`, exactly as the
+// search index already does. That is deliberate, because the adversarial review
+// of the original full-replica design killed four of its findings by removing the
+// worker rather than fixing them:
+//
+//   C2 - "credentials are request-scoped, so no persistent worker can hold them".
+//        Still true, and still fine, because there is no worker. A cycle only ever
+//        reads an ALREADY-MINTED Authorization header off the request.
+//   C3 - the OAuth-refresh mitigation being itself the bug. Avoided by
+//        construction: nothing here touches the refresh-token cookie, so it cannot
+//        rotate a token into a response nobody reads and log the user out.
+//   C4 - a shared `registry.json` breaking the epoch fencing token. No registry,
+//        no epochs: single-flight per account inside one process, and every piece
+//        of state lives in the SQLite file under a real transaction.
+//   H1 - "a server-side engine cannot read a renderer-only setting". The renderer
+//        decides when to sync, so nothing materialises for an account that never
+//        opted in. The retention POLICY does need to be durable server-side, so it
+//        lives inside the encrypted store (written through PUT /api/offline/status),
+//        not in renderer localStorage.
+//   H4 - unbounded concurrent multi-account sync starving foreground activity.
+//        One request, one account, one cycle, hard budgets below.
+//
+// H2 (key handoff via process environment) was already fixed by what shipped: the
+// key crosses on an inherited file descriptor and is zeroed after each job.
+// C1 (the native dependency breaking Alpine `docker build`s) is likewise already
+// fixed and this file adds no new dependency.
+//
+// H3 is the one finding that genuinely COMES BACK. The webmail does local delta
+// arithmetic on mailbox unread counts and totals for mark-read/move/delete, and a
+// read-only offline cache sitting underneath that arithmetic needs a coherence
+// story. The story is in `read.ts`: the replica is a FALLBACK, never a cache in
+// front of the server. It is consulted only after a read has actually failed at
+// the transport level, so an online session never sees a replica count.
+//
+// JOB ORDER within a cycle is A1 (Mailbox/changes) -> A2 (Email/changes) ->
+// B (coverage enumeration) -> C1 (body queue drain) -> C2 (body backfill) ->
+// retention. The three machines are logically independent but OPERATIONALLY
+// SERIALISED. Either job order is safe; CONCURRENCY is not - coverage's
+// query-then-apply pair interleaved with the delta path's apply resurrects a
+// destroyed message as a zombie no future `/changes` page will ever re-report.
+// "Run bodies in parallel, it's separate state" is forbidden for the same reason.
+
+import type { Email, Mailbox } from '@/lib/jmap/types';
+import { logger } from '@/lib/logger';
+import type { JmapSessionInfo } from '@/lib/mail-index/jmap';
+import {
+  madeForwardProgress, planEmailFetches, planMailboxFetches, pageIsEmpty, advanceOneMs,
+} from './apply';
+import {
+  backoffDelayMs, escalationApplies, movesCursor, nextRung, ReplicaSyncError, rungValue,
+} from './errors';
+import {
+  AnchorNotFoundError, captureSnapshotStates, getEmailChanges, getEmails, getMailboxChanges,
+  getMailboxes, maxObjectsInGet, queryAscending,
+} from './jmap';
+import {
+  adjustForWindow, computeFloors, floorMovement, guardFloorAgainstClockJump,
+} from './retention';
+import { CURSOR_TYPES, mintEnumerationCommitment, type CursorType } from './states';
+import { ReplicaStore, type RetentionPolicy } from './store';
+import type { BodyQueueEntry, CoverageState, EnvelopeRow, MailboxRow } from './types';
+
+// ── budgets ──────────────────────────────────────────────────────────────────
+// A cycle runs inside an HTTP request the renderer is waiting on, so the wall
+// clock matters more than page counts. Budget exhaustion is a NORMAL outcome
+// reported as `unfinishedWork`, never an error.
+export const BUDGET = {
+  changesPagesPerCursor: 20,
+  coveragePages: 12,
+  coveragePageSize: 200,
+  bodyItems: 60,
+  bodyFetchChunk: 10,
+  envelopeFetchChunk: 50,
+  wallClockMs: 45_000,
+} as const;
+
+export const MAX_BODY_ATTEMPTS = 5;
+/** Reconciles per rolling 24 h before throttling. Never a hard stop. */
+export const MAX_RECONCILES_PER_DAY = 4;
+
+/**
+ * The reconcile stamp, derived from the DATA and not from the clock.
+ *
+ * Exported so it is testable on its own, because getting it wrong fails SILENTLY
+ * in both directions. With a frozen or coarse clock, a plain `now` leaves
+ * `cached_at < stamp` matching nothing and the sweep deletes nothing at all. And
+ * because the pin routinely EXCEEDS `now`, any enumeration path that stamps with
+ * `now` instead of the pin leaves its rows below the pin and gets them deleted by
+ * the very sweep that just re-verified them against the server.
+ */
+export function reconcileStamp(now: number, maxCachedAt: number): number {
+  return Math.max(now, maxCachedAt + 1);
+}
+
+export interface CycleContext {
+  store: ReplicaStore;
+  session: JmapSessionInfo;
+  authHeader: string;
+  jmapAccountId: string;
+  policy: RetentionPolicy;
+  now: number;
+  deadline: number;
+}
+
+export interface CycleReport {
+  ok: boolean;
+  /** True when budgets ran out or a queue still has wanted work - chain another cycle. */
+  unfinishedWork: boolean;
+  bootstrapped: boolean;
+  reconciled: boolean;
+  mailboxesWritten: number;
+  envelopesWritten: number;
+  envelopesDeleted: number;
+  bodiesWritten: number;
+  bodiesEvicted: number;
+  coveragePhase: CoverageState['phase'];
+  resyncRequired: boolean;
+  warnings: string[];
+  errorClass?: string;
+  error?: string;
+  retryAfterMs?: number;
+  durationMs: number;
+}
+
+// ── row conversion ───────────────────────────────────────────────────────────
+
+function toEnvelopeRow(jmapAccountId: string, email: Email): EnvelopeRow | null {
+  // `received_at` is NOT NULL and drives every window, so a record without one
+  // cannot be stored. In practice the server always sends it for a full envelope
+  // fetch; skipping is the safe response to a malformed one.
+  if (typeof email.receivedAt !== 'string' || email.receivedAt.length === 0) return null;
+  return {
+    jmapAccountId,
+    id: email.id,
+    threadId: email.threadId ?? null,
+    receivedAt: email.receivedAt,
+    size: typeof email.size === 'number' ? email.size : null,
+    subject: email.subject ?? null,
+    preview: email.preview ?? null,
+    fromJson: email.from ? JSON.stringify(email.from) : null,
+    toJson: email.to ? JSON.stringify(email.to) : null,
+    ccJson: email.cc ? JSON.stringify(email.cc) : null,
+    blobId: email.blobId ?? null,
+    hasAttachment: email.hasAttachment === true,
+    keywordsJson: JSON.stringify(email.keywords ?? {}),
+    // The BARE JMAP mailbox ids, never the display layer's `<accountId>:<id>`
+    // prefixed form. The store already keys every row by (jmapAccountId, id), so
+    // a prefixed id would double-encode the account and break every lookup.
+    mailboxIds: Object.entries(email.mailboxIds ?? {})
+      .filter(([, v]) => v)
+      .map(([k]) => k),
+  };
+}
+
+function toMailboxRow(jmapAccountId: string, mailbox: Mailbox): MailboxRow {
+  return {
+    jmapAccountId,
+    id: mailbox.originalId ?? mailbox.id,
+    name: mailbox.name ?? '',
+    parentId: mailbox.parentId ?? null,
+    role: mailbox.role ?? null,
+    sortOrder: typeof mailbox.sortOrder === 'number' ? mailbox.sortOrder : null,
+    totalEmails: typeof mailbox.totalEmails === 'number' ? mailbox.totalEmails : null,
+    unreadEmails: typeof mailbox.unreadEmails === 'number' ? mailbox.unreadEmails : null,
+    totalThreads: typeof mailbox.totalThreads === 'number' ? mailbox.totalThreads : null,
+    unreadThreads: typeof mailbox.unreadThreads === 'number' ? mailbox.unreadThreads : null,
+    myRightsJson: mailbox.myRights ? JSON.stringify(mailbox.myRights) : null,
+    isSubscribed: mailbox.isSubscribed !== false,
+  };
+}
+
+/** The body tier as one opaque JSON blob. */
+export function serialiseBody(email: Email): string {
+  return JSON.stringify({
+    sentAt: email.sentAt,
+    bcc: email.bcc,
+    replyTo: email.replyTo,
+    textBody: email.textBody,
+    htmlBody: email.htmlBody,
+    bodyValues: email.bodyValues,
+    attachments: email.attachments,
+    messageId: email.messageId,
+    inReplyTo: email.inReplyTo,
+    references: email.references,
+    headers: email.headers,
+    bodyStructure: email.bodyStructure,
+  });
+}
+
+// ── the cycle ────────────────────────────────────────────────────────────────
+
+export async function runCycle(ctx: CycleContext): Promise<CycleReport> {
+  const started = Date.now();
+  const report: CycleReport = {
+    ok: true,
+    unfinishedWork: false,
+    bootstrapped: false,
+    reconciled: false,
+    mailboxesWritten: 0,
+    envelopesWritten: 0,
+    envelopesDeleted: 0,
+    bodiesWritten: 0,
+    bodiesEvicted: 0,
+    coveragePhase: 'never-run',
+    resyncRequired: false,
+    warnings: [],
+    durationMs: 0,
+  };
+
+  const { store, jmapAccountId, now } = ctx;
+  const flags = store.getFlags(now);
+  const rawFloors = computeFloors(ctx.policy, now);
+
+  // Intent vs glitch. Without this discriminator the clock guard also fires on a
+  // legitimate user retention change, leaving a Settings edit unapplied until
+  // some unrelated trigger happened to move the floor again.
+  const policyChanged =
+    flags.lastEnvelopeDays !== undefined && flags.lastEnvelopeDays !== ctx.policy.envelopeDays;
+  const guarded = guardFloorAgainstClockJump(rawFloors.envelopeFrom, flags.lastWindowFloor, {
+    policyChanged,
+  });
+  if (guarded.warning) {
+    report.warnings.push(guarded.warning);
+    logger.warn('offline-replica: retention floor suppressed', { warning: guarded.warning });
+  }
+  const envelopeFrom = guarded.envelopeFrom;
+  // Clamp a second time after the guard: a suppressed envelope floor can end up
+  // NEWER than the computed body floor, and a body without an envelope is an orphan.
+  const bodyFrom = rawFloors.bodyFrom > envelopeFrom ? rawFloors.bodyFrom : envelopeFrom;
+
+  try {
+    let coverage = store.getCoverage(jmapAccountId);
+
+    // ── bootstrap / reconcile ────────────────────────────────────────────────
+    if (!coverage || coverage.phase === 'never-run') {
+      await beginEnumeration(ctx, envelopeFrom, 'bootstrap');
+      report.bootstrapped = true;
+      coverage = store.getCoverage(jmapAccountId);
+    } else if (flags.resyncRequired && coverage.phase !== 'reconciling') {
+      if (reconcileBudgetAllows(store, flags, now)) {
+        await beginEnumeration(ctx, envelopeFrom, 'reconcile');
+        report.reconciled = true;
+        coverage = store.getCoverage(jmapAccountId);
+      } else {
+        // Throttle, never stop: a hard stop would trade a reconcile loop for a
+        // permanent wedge.
+        report.warnings.push('reconcile throttled: more than 4 rebuilds in the last 24h');
+      }
+    }
+
+    // ── A1 / A2: drain the two /changes cursors ──────────────────────────────
+    let anyCursorPending = false;
+    for (const type of CURSOR_TYPES) {
+      if (Date.now() > ctx.deadline) { report.unfinishedWork = true; break; }
+      const drained = await drainCursor(ctx, type, bodyFrom, report);
+      if (drained.pending) anyCursorPending = true;
+    }
+    if (anyCursorPending) report.unfinishedWork = true;
+
+    // ── B: coverage enumeration ──────────────────────────────────────────────
+    coverage = store.getCoverage(jmapAccountId);
+    if (coverage && (coverage.phase === 'scanning' || coverage.phase === 'reconciling')) {
+      const scanned = await runCoverage(ctx, coverage, bodyFrom, guarded.evictionAllowed, report);
+      if (scanned.unfinished) report.unfinishedWork = true;
+    }
+
+    // ── C1 / C2: bodies ──────────────────────────────────────────────────────
+    const drainedBodies = await drainBodyQueue(ctx, report);
+    const backfilled = await backfillBodies(ctx, bodyFrom, rawFloors.maxBodyBytes, drainedBodies);
+    if (backfilled > 0) report.unfinishedWork = true;
+    if (store.countWantedBodies(jmapAccountId, Date.now()) > 0) report.unfinishedWork = true;
+
+    // ── retention ────────────────────────────────────────────────────────────
+    applyRetention(ctx, {
+      envelopeFrom,
+      bodyFrom,
+      maxBodyBytes: rawFloors.maxBodyBytes,
+      evictionAllowed: guarded.evictionAllowed,
+      previousFloor: flags.lastWindowFloor,
+      report,
+    });
+
+    const finalCoverage = store.getCoverage(jmapAccountId);
+    const finalFlags = store.getFlags(now);
+    report.coveragePhase = finalCoverage?.phase ?? 'never-run';
+    report.resyncRequired = finalFlags.resyncRequired;
+
+    store.transaction(() => {
+      store.patchFlags(now, {
+        // The USED floor, never the rejected one.
+        lastWindowFloor: guarded.nextLastWindowFloor,
+        lastEnvelopeDays: ctx.policy.envelopeDays,
+        lastMaxBodyBytes: rawFloors.maxBodyBytes,
+        lastCycleAt: now,
+        lastCycleOk: true,
+        lastCycleError: undefined,
+      });
+    });
+  } catch (error) {
+    report.ok = false;
+    if (error instanceof ReplicaSyncError) {
+      report.errorClass = error.cls;
+      report.error = error.message;
+      report.retryAfterMs = error.retryAfterMs;
+      // "Offline is not an error" - a transport failure leaves every cursor
+      // exactly where it was and is simply retried later.
+      report.unfinishedWork = error.cls !== 'Fatal' && error.cls !== 'Auth';
+    } else {
+      report.error = error instanceof Error ? error.message : String(error);
+      report.unfinishedWork = false;
+    }
+    try {
+      store.transaction(() => {
+        store.patchFlags(now, { lastCycleAt: now, lastCycleOk: false, lastCycleError: report.error });
+      });
+    } catch { /* the store may be the thing that failed */ }
+  }
+
+  report.durationMs = Date.now() - started;
+  return report;
+}
+
+// ── bootstrap / reconcile ────────────────────────────────────────────────────
+
+function reconcileBudgetAllows(
+  store: ReplicaStore,
+  flags: { reconcilesInWindow: number; reconcileWindowStartedAt: number },
+  now: number,
+): boolean {
+  const dayMs = 24 * 60 * 60 * 1000;
+  if (now - flags.reconcileWindowStartedAt > dayMs) {
+    store.transaction(() => {
+      store.patchFlags(now, { reconcilesInWindow: 0, reconcileWindowStartedAt: now });
+    });
+    return true;
+  }
+  return flags.reconcilesInWindow < MAX_RECONCILES_PER_DAY;
+}
+
+/**
+ * THE MANDATORY ORDER. Step 1 must precede step 3.
+ *
+ *  1. capture both cursors, in one request, BEFORE touching any data, and seed
+ *     them inside one EnumerationCommitment that writes the coverage row in the
+ *     SAME transaction - so a seed is never durable without the durable promise to
+ *     enumerate that justifies it.
+ *  2. full `Mailbox/get` - cheap, always complete, no paging.
+ *  3. the seeded cursors are LIVE FROM HERE. Each cycle runs the delta jobs and
+ *     only then the scan, so a wide-window rebuild does not stall incoming mail.
+ *  4. when the scan reaches the target: `coveredFrom = sweepFloor`, phase complete.
+ */
+async function beginEnumeration(
+  ctx: CycleContext,
+  envelopeFrom: string,
+  kind: 'bootstrap' | 'reconcile',
+): Promise<void> {
+  const { store, jmapAccountId, now } = ctx;
+
+  // Step 0 for a reconcile: PIN THE FLOOR. Every later step reads `sweepFloor`,
+  // never a live `targetFrom`. Widening retention while a reconcile runs would
+  // otherwise make the sweep delete against the new wide window when the
+  // enumeration only covered the old narrow one - permanently, since `coveredFrom`
+  // then claims the wider range and `/changes` cannot re-deliver old mail.
+  const existing = store.getCoverage(jmapAccountId);
+  const sweepFloor = kind === 'reconcile' ? (existing?.deferredTargetFrom ?? envelopeFrom) : envelopeFrom;
+
+  const snapshots = await captureSnapshotStates(ctx.session, ctx.authHeader, jmapAccountId);
+
+  // Derive the reconcile stamp from the DATA, not the clock: with a frozen or
+  // coarse clock `cached_at < stamp` matches nothing and the sweep deletes nothing.
+  const stampedAt = kind === 'reconcile'
+    ? reconcileStamp(now, store.maxEnvelopeCachedAt(jmapAccountId))
+    : undefined;
+
+  store.transaction(() => {
+    for (const type of CURSOR_TYPES) {
+      store.seedCursor(
+        { jmapAccountId, type },
+        mintEnumerationCommitment({
+          jmapAccountId,
+          snapshot: type === 'Mailbox' ? snapshots.mailbox : snapshots.email,
+          targetFrom: envelopeFrom,
+          sweepFloor,
+          kind,
+        }),
+        now,
+      );
+    }
+    if (stampedAt !== undefined) store.patchCoverage(jmapAccountId, { reconcileStampedAt: stampedAt });
+    if (kind === 'reconcile') {
+      const flags = store.getFlags(now);
+      store.patchFlags(now, { reconcilesInWindow: flags.reconcilesInWindow + 1 });
+    }
+  });
+
+  // Step 2: every mailbox, in full.
+  const mailboxes = await getMailboxes(ctx.session, ctx.authHeader, jmapAccountId, null);
+  const rows = mailboxes.map((m) => toMailboxRow(jmapAccountId, m));
+  store.transaction(() => { store.upsertMailboxes(rows); });
+}
+
+// ── A1 / A2: the delta drain ─────────────────────────────────────────────────
+
+async function drainCursor(
+  ctx: CycleContext,
+  type: CursorType,
+  bodyFrom: string,
+  report: CycleReport,
+): Promise<{ pending: boolean }> {
+  const { store, jmapAccountId } = ctx;
+  const key = { jmapAccountId, type };
+  let cursor = store.getCursor(key);
+  if (!cursor) return { pending: false };
+  if (cursor.invalidatedAt) {
+    // Serving `/changes` from an invalidated cursor is forbidden. The reconcile
+    // is what clears it.
+    return { pending: false };
+  }
+
+  const cap = maxObjectsInGet(ctx.session);
+  let pages = 0;
+
+  while (pages < BUDGET.changesPagesPerCursor) {
+    if (Date.now() > ctx.deadline) {
+      store.transaction(() => { store.patchCursor(key, { drainPending: true }); });
+      return { pending: true };
+    }
+    cursor = store.getCursor(key);
+    if (!cursor) return { pending: false };
+
+    let page;
+    try {
+      const maxChanges = rungValue(cursor.maxChangesRung, cap);
+      page = type === 'Email'
+        ? await getEmailChanges(ctx.session, ctx.authHeader, jmapAccountId, cursor.state, maxChanges)
+        : await getMailboxChanges(ctx.session, ctx.authHeader, jmapAccountId, cursor.state, maxChanges);
+    } catch (error) {
+      handleDrainError(ctx, key, cursor.state, error, report);
+      // Any cursor failing means the cycle is unfinished for escalation
+      // purposes, but never that the cursor moved.
+      return { pending: true };
+    }
+    pages++;
+
+    if (page.oldState !== cursor.state) {
+      // Re-issue once before escalating: a transient anomaly is far more common
+      // than a genuine invalidation, and a full rebuild is expensive.
+      let reissued;
+      try {
+        const maxChanges = rungValue(cursor.maxChangesRung, cap);
+        reissued = type === 'Email'
+          ? await getEmailChanges(ctx.session, ctx.authHeader, jmapAccountId, cursor.state, maxChanges)
+          : await getMailboxChanges(ctx.session, ctx.authHeader, jmapAccountId, cursor.state, maxChanges);
+      } catch (error) {
+        handleDrainError(ctx, key, cursor.state, error, report);
+        return { pending: true };
+      }
+      if (reissued.oldState !== cursor.state) {
+        invalidate(ctx, key, 'oldStateMismatch', report);
+        return { pending: false };
+      }
+      // Use the RE-ISSUED page. Advancing to the original response's newState
+      // would skip whatever the re-issue reported - the same silent-gap shape as
+      // the cursor-provenance bug, reintroduced one level down.
+      page = reissued;
+      report.warnings.push(`${type}/changes reported a transient oldState mismatch`);
+    }
+
+    if (pageIsEmpty(page)) {
+      // An empty page still advances. Skipping it re-requests forever.
+      store.transaction(() => {
+        store.advanceCursor(key, page.newState);
+        store.patchCursor(key, { consecutiveFailures: 0, lastFailedState: undefined, maxChangesRung: 0 });
+      });
+      if (!page.hasMoreChanges) {
+        store.transaction(() => { store.patchCursor(key, { drainPending: false }); });
+        return { pending: false };
+      }
+      continue;
+    }
+
+    if (type === 'Mailbox') {
+      await applyMailboxPage(ctx, page, report);
+    } else {
+      await applyEmailPage(ctx, page, bodyFrom, report);
+    }
+
+    store.transaction(() => {
+      store.advanceCursor(key, page.newState);
+      store.patchCursor(key, {
+        consecutiveFailures: 0,
+        lastFailedState: undefined,
+        maxChangesRung: 0,
+        drainPending: page.hasMoreChanges,
+      });
+    });
+
+    if (!page.hasMoreChanges) return { pending: false };
+  }
+
+  // Budget exhaustion is normal, not an error. It is also the answer to a server
+  // whose `hasMoreChanges` never goes false.
+  store.transaction(() => { store.patchCursor(key, { drainPending: true }); });
+  return { pending: true };
+}
+
+function handleDrainError(
+  ctx: CycleContext,
+  key: { jmapAccountId: string; type: CursorType },
+  failedState: string,
+  error: unknown,
+  report: CycleReport,
+): void {
+  const { store } = ctx;
+  const cls = error instanceof ReplicaSyncError ? error.cls : 'ServerTransient';
+  const message = error instanceof Error ? error.message : String(error);
+  report.warnings.push(`${key.type}/changes ${cls}: ${message}`);
+  if (error instanceof ReplicaSyncError && error.retryAfterMs) {
+    report.retryAfterMs = Math.max(report.retryAfterMs ?? 0, error.retryAfterMs);
+  }
+
+  if (movesCursor(cls)) {
+    invalidate(ctx, key, 'cannotCalculateChanges', report);
+    return;
+  }
+
+  const cursor = store.getCursor(key);
+  if (!cursor) return;
+  // The ladder only counts failures at the SAME sinceState: a failure at a new
+  // position means progress was made, so the ladder restarts.
+  const sameSpot = cursor.lastFailedState === failedState;
+  const failures = sameSpot ? cursor.consecutiveFailures + 1 : 1;
+  store.transaction(() => {
+    store.patchCursor(key, {
+      consecutiveFailures: failures,
+      lastFailedState: failedState,
+      maxChangesRung: escalationApplies(cls) && failures >= 2
+        ? nextRung(cursor.maxChangesRung)
+        : cursor.maxChangesRung,
+    });
+  });
+}
+
+/**
+ * RFC 8620 s5.2 says the client MUST invalidate its cache. The LITERAL reading -
+ * delete everything, now - would empty a user's offline mail exactly when they may
+ * be offline and depending on it. So: mark the cursor unusable, set the STICKY
+ * resync flag, and leave every record READABLE. The accepted cost is that between
+ * detection and the sweep, a server-deleted message can still show locally.
+ *
+ * An invalidation of EITHER cursor reconciles the account as a whole - splitting
+ * it is not worth the reasoning burden when `Mailbox/get` is one cheap call.
+ */
+function invalidate(
+  ctx: CycleContext,
+  key: { jmapAccountId: string; type: CursorType },
+  reason: 'cannotCalculateChanges' | 'oldStateMismatch',
+  report: CycleReport,
+): void {
+  const { store, now } = ctx;
+  store.transaction(() => {
+    store.patchCursor(key, { invalidatedAt: now, invalidatedReason: reason });
+    store.patchFlags(now, { resyncRequired: true });
+  });
+  report.resyncRequired = true;
+  report.warnings.push(`${key.type} cursor invalidated (${reason}); a rebuild is queued`);
+  logger.warn('offline-replica: cursor invalidated', { type: key.type, reason });
+}
+
+async function applyMailboxPage(
+  ctx: CycleContext,
+  page: Parameters<typeof planMailboxFetches>[0],
+  report: CycleReport,
+): Promise<void> {
+  const { store, jmapAccountId } = ctx;
+  const plan = planMailboxFetches(page);
+
+  if (plan.fullIds.length > 0) {
+    const mailboxes = await getMailboxes(ctx.session, ctx.authHeader, jmapAccountId, plan.fullIds);
+    const rows = mailboxes.map((m) => toMailboxRow(jmapAccountId, m));
+    store.transaction(() => { report.mailboxesWritten += store.upsertMailboxes(rows); });
+  }
+
+  if (plan.countOnlyIds.length > 0) {
+    const props = ['totalEmails', 'unreadEmails', 'totalThreads', 'unreadThreads'];
+    const mailboxes = await getMailboxes(
+      ctx.session, ctx.authHeader, jmapAccountId, plan.countOnlyIds, props,
+    );
+    store.transaction(() => {
+      for (const m of mailboxes) {
+        store.patchMailboxCounts(jmapAccountId, m.originalId ?? m.id, {
+          totalEmails: typeof m.totalEmails === 'number' ? m.totalEmails : undefined,
+          unreadEmails: typeof m.unreadEmails === 'number' ? m.unreadEmails : undefined,
+          totalThreads: typeof m.totalThreads === 'number' ? m.totalThreads : undefined,
+          unreadThreads: typeof m.unreadThreads === 'number' ? m.unreadThreads : undefined,
+        });
+      }
+    });
+  }
+
+  if (plan.destroyIds.length > 0) {
+    // The mailbox row ONLY. Never email records.
+    store.transaction(() => { store.deleteMailboxes(jmapAccountId, plan.destroyIds); });
+  }
+}
+
+async function applyEmailPage(
+  ctx: CycleContext,
+  page: Parameters<typeof planEmailFetches>[0],
+  bodyFrom: string,
+  report: CycleReport,
+): Promise<void> {
+  const { store, jmapAccountId, now } = ctx;
+  // Presence is tested in BULK, before either fetch is issued - an `updated` id we
+  // do not hold is filtered out rather than fetched and then discarded.
+  const present = store.whichEnvelopesExist(jmapAccountId, page.updated);
+  const plan = planEmailFetches(page, present);
+
+  // CREATES: full envelope tier, and a body enqueue when inside the body window.
+  // Bodies are NEVER fetched inline - they are 10-500 KB against an envelope's
+  // ~1 KB, and the queue is what keeps a page a small, quickly-committable unit.
+  for (let i = 0; i < plan.createIds.length; i += BUDGET.envelopeFetchChunk) {
+    const chunk = plan.createIds.slice(i, i + BUDGET.envelopeFetchChunk);
+    const { list } = await getEmails(ctx.session, ctx.authHeader, jmapAccountId, chunk, 'envelope');
+    const rows = list.map((e) => toEnvelopeRow(jmapAccountId, e)).filter((r): r is EnvelopeRow => r !== null);
+    const queue: BodyQueueEntry[] = rows
+      .filter((r) => r.receivedAt >= bodyFrom)
+      .map((r) => ({ emailId: r.id, jmapAccountId, receivedAt: r.receivedAt, attempts: 0 }));
+    store.transaction(() => {
+      report.envelopesWritten += store.upsertEnvelopes(rows, now);
+      store.enqueueBodies(queue);
+    });
+  }
+
+  // UPDATES: three properties, never a body.
+  for (let i = 0; i < plan.updateIds.length; i += BUDGET.envelopeFetchChunk) {
+    const chunk = plan.updateIds.slice(i, i + BUDGET.envelopeFetchChunk);
+    const { list } = await getEmails(ctx.session, ctx.authHeader, jmapAccountId, chunk, 'mutable');
+    store.transaction(() => {
+      for (const e of list) {
+        store.patchEnvelopeMutable(jmapAccountId, e.id, {
+          keywordsJson: JSON.stringify(e.keywords ?? {}),
+          mailboxIds: Object.entries(e.mailboxIds ?? {}).filter(([, v]) => v).map(([k]) => k),
+        });
+      }
+    });
+  }
+
+  // DESTROYS LAST. Ids are never reused, so a destroy always refers to the same
+  // record as any create/update of that id in the same page, and destroy-last
+  // converges. The reverse order would resurrect a dead id, spend a fetch and get
+  // `notFound`. A destroy for an id we never held is a harmless no-op.
+  if (plan.destroyIds.length > 0) {
+    store.transaction(() => {
+      report.envelopesDeleted += store.deleteEmails(jmapAccountId, plan.destroyIds);
+    });
+  }
+}
+
+// ── B: coverage enumeration ──────────────────────────────────────────────────
+
+async function runCoverage(
+  ctx: CycleContext,
+  coverage: CoverageState,
+  bodyFrom: string,
+  evictionAllowed: boolean,
+  report: CycleReport,
+): Promise<{ unfinished: boolean }> {
+  const { store, jmapAccountId, now } = ctx;
+  const isReconcile = coverage.phase === 'reconciling';
+  const floor = isReconcile ? (coverage.sweepFloor ?? coverage.targetFrom) : coverage.targetFrom;
+  // A reconcile stamps with its PINNED value, not `now`. The pin is
+  // max(now, maxCachedAt + 1) so it routinely EXCEEDS now - stamping with `now`
+  // would leave rows below the pin and get them deleted by the very sweep that
+  // just re-verified them against the server.
+  const stamp = isReconcile ? coverage.reconcileStampedAt ?? now : now;
+
+  let scanCursor = coverage.scanCursor ?? floor;
+  let lastPageIds: string[] = [];
+  let pages = 0;
+
+  while (pages < BUDGET.coveragePages) {
+    if (Date.now() > ctx.deadline) return { unfinished: true };
+    pages++;
+
+    let ids: string[];
+    try {
+      ({ ids } = await queryAscending(
+        ctx.session, ctx.authHeader, jmapAccountId, scanCursor, BUDGET.coveragePageSize,
+      ));
+    } catch (error) {
+      if (error instanceof AnchorNotFoundError) {
+        ids = [];
+      } else {
+        store.transaction(() => {
+          store.patchCoverage(jmapAccountId, {
+            consecutiveFailures: coverage.consecutiveFailures + 1,
+          });
+        });
+        throw error;
+      }
+    }
+
+    if (ids.length === 0) {
+      // The scan has reached the present.
+      finishEnumeration(ctx, floor, stamp, isReconcile, evictionAllowed, report);
+      return { unfinished: !evictionAllowed && isReconcile };
+    }
+
+    const { list } = await getEmails(ctx.session, ctx.authHeader, jmapAccountId, ids, 'envelope');
+    const rows = list.map((e) => toEnvelopeRow(jmapAccountId, e)).filter((r): r is EnvelopeRow => r !== null);
+    const maxReceivedAt = rows.reduce<string | null>(
+      (max, r) => (max === null || r.receivedAt > max ? r.receivedAt : max),
+      null,
+    );
+    const queue: BodyQueueEntry[] = rows
+      .filter((r) => r.receivedAt >= bodyFrom)
+      .map((r) => ({ emailId: r.id, jmapAccountId, receivedAt: r.receivedAt, attempts: 0 }));
+
+    let nextScanCursor = scanCursor;
+    let gapMarker: NonNullable<CoverageState['gapMarkers']>[number] | null = null;
+
+    if (madeForwardProgress(maxReceivedAt, scanCursor)) {
+      nextScanCursor = maxReceivedAt as string;
+    } else if (ids.length >= BUDGET.coveragePageSize) {
+      // A FULL page whose every row shares one millisecond. Try the anchor rung
+      // first: resume from the id after the last one we saw.
+      const anchorId = lastPageIds.length > 0 ? lastPageIds[lastPageIds.length - 1] : ids[ids.length - 1];
+      let recovered = false;
+      try {
+        const anchored = await queryAscending(
+          ctx.session, ctx.authHeader, jmapAccountId, scanCursor, BUDGET.coveragePageSize,
+          { anchor: anchorId, anchorOffset: 1 },
+        );
+        if (anchored.ids.length > 0) {
+          const fetched = await getEmails(
+            ctx.session, ctx.authHeader, jmapAccountId, anchored.ids, 'envelope',
+          );
+          const anchoredRows = fetched.list
+            .map((e) => toEnvelopeRow(jmapAccountId, e))
+            .filter((r): r is EnvelopeRow => r !== null);
+          const anchoredMax = anchoredRows.reduce<string | null>(
+            (max, r) => (max === null || r.receivedAt > max ? r.receivedAt : max),
+            null,
+          );
+          store.transaction(() => {
+            // The PINNED stamp here too, for the same reason.
+            report.envelopesWritten += store.upsertEnvelopes(anchoredRows, stamp);
+            store.enqueueBodies(
+              anchoredRows
+                .filter((r) => r.receivedAt >= bodyFrom)
+                .map((r) => ({ emailId: r.id, jmapAccountId, receivedAt: r.receivedAt, attempts: 0 })),
+            );
+          });
+          if (anchoredMax !== null && anchoredMax > scanCursor) {
+            nextScanCursor = anchoredMax;
+            recovered = true;
+          }
+        }
+      } catch (error) {
+        if (!(error instanceof AnchorNotFoundError)) throw error;
+      }
+      if (!recovered) {
+        // Last resort: advance one millisecond, WARN, and leave a durable trace.
+        // This rung CAN skip messages sharing the boundary millisecond, so it is
+        // never normal-path behaviour and always leaves a record so a support
+        // question has an answer. A 200-message single-millisecond cluster is a
+        // corrupt server, not a case to design for.
+        const to = advanceOneMs(scanCursor);
+        gapMarker = { from: scanCursor, to, reason: 'tie-cluster-skip', at: now };
+        nextScanCursor = to;
+        report.warnings.push(
+          `coverage skipped a tie cluster at ${scanCursor}; some messages sharing that ` +
+            `millisecond may be missing from the offline store`,
+        );
+        logger.warn('offline-replica: tie-cluster skip', { at: scanCursor });
+      }
+    } else {
+      // A partial page with no forward progress means we are at the tail.
+      //
+      // COMMIT THE ROWS BEFORE FINISHING. Finishing runs the reconcile sweep, and
+      // the sweep deletes anything still stamped below the pin - so finishing
+      // first would delete this page's own records (taking their bodies and queue
+      // rows with them) and then re-insert them bodyless, costing a re-download of
+      // every body at the tail on every reconcile.
+      store.transaction(() => {
+        report.envelopesWritten += store.upsertEnvelopes(rows, stamp);
+        store.enqueueBodies(queue);
+      });
+      finishEnumeration(ctx, floor, stamp, isReconcile, evictionAllowed, report);
+      return { unfinished: false };
+    }
+
+    const advanceTo = nextScanCursor;
+    const marker = gapMarker;
+    store.transaction(() => {
+      report.envelopesWritten += store.upsertEnvelopes(rows, stamp);
+      store.enqueueBodies(queue);
+      // CURSOR LAST, inside the same transaction as the records it accounts for.
+      store.patchCoverage(jmapAccountId, {
+        scanCursor: advanceTo,
+        seen: coverage.seen + rows.length,
+        consecutiveFailures: 0,
+        ...(marker
+          ? { gapMarkers: [...(coverage.gapMarkers ?? []), marker].slice(-32) }
+          : {}),
+      });
+    });
+    scanCursor = advanceTo;
+    lastPageIds = ids;
+  }
+
+  return { unfinished: true };
+}
+
+/**
+ * The only place the reconcile pins are released, and it clears them all at once.
+ */
+function finishEnumeration(
+  ctx: CycleContext,
+  floor: string,
+  stamp: number,
+  isReconcile: boolean,
+  evictionAllowed: boolean,
+  report: CycleReport,
+): void {
+  const { store, jmapAccountId, now } = ctx;
+  const coverage = store.getCoverage(jmapAccountId);
+  if (!coverage) return;
+
+  if (isReconcile && !evictionAllowed) {
+    // The sweep is a delete against the retention floor, so it is subject to the
+    // same rule as eviction: a suspect clock reading may not drive deletion.
+    // Leave the reconcile open; it completes on a cycle whose floor is trustworthy.
+    report.warnings.push(
+      'deferring the reconcile sweep: the retention floor came from a suppressed clock anomaly',
+    );
+    return;
+  }
+
+  store.transaction(() => {
+    if (isReconcile) {
+      report.envelopesDeleted += store.sweep(jmapAccountId, floor, stamp);
+      // A give-up recorded during whatever went wrong must not outlive it, or a
+      // transient outage would permanently deny a body with no path back.
+      store.clearBodyGiveUps(jmapAccountId);
+      store.patchFlags(now, { resyncRequired: false });
+      for (const type of CURSOR_TYPES) {
+        store.patchCursor({ jmapAccountId, type }, {
+          invalidatedAt: undefined,
+          invalidatedReason: undefined,
+          consecutiveFailures: 0,
+          lastFailedState: undefined,
+          maxChangesRung: 0,
+        });
+      }
+    }
+    const deferred = coverage.deferredTargetFrom;
+    store.patchCoverage(jmapAccountId, {
+      coveredFrom: floor,
+      scanCursor: null,
+      sweepFloor: undefined,
+      reconcileStampedAt: undefined,
+      deferredTargetFrom: undefined,
+      ...(deferred ? { targetFrom: deferred, phase: 'scanning' as const } : { phase: 'complete' as const }),
+    });
+  });
+}
+
+// ── C1: drain the durable body queue ─────────────────────────────────────────
+
+async function drainBodyQueue(ctx: CycleContext, report: CycleReport): Promise<number> {
+  const { store, jmapAccountId } = ctx;
+  let fetched = 0;
+  const wanted = store.takeBodyQueue(jmapAccountId, BUDGET.bodyItems, Date.now());
+  if (wanted.length === 0) return 0;
+
+  for (let i = 0; i < wanted.length; i += BUDGET.bodyFetchChunk) {
+    if (Date.now() > ctx.deadline) { report.unfinishedWork = true; break; }
+    const chunk = wanted.slice(i, i + BUDGET.bodyFetchChunk);
+    let result;
+    try {
+      result = await getEmails(
+        ctx.session, ctx.authHeader, jmapAccountId, chunk.map((e) => e.emailId), 'body',
+      );
+    } catch (error) {
+      // Body failures NEVER touch a cursor: the body jobs are separate state, so
+      // the delta path keeps its position and the queue simply retries later.
+      const message = error instanceof Error ? error.message : String(error);
+      store.transaction(() => {
+        for (const e of chunk) {
+          store.bumpBodyAttempt(jmapAccountId, e.emailId, Date.now() + backoffDelayMs(e.attempts), message);
+        }
+      });
+      report.warnings.push(`body fetch failed: ${message}`);
+      report.unfinishedWork = true;
+      break;
+    }
+
+    const byId = new Map(result.list.map((e) => [e.id, e]));
+    const notFound = new Set(result.notFound);
+
+    store.transaction(() => {
+      for (const entry of chunk) {
+        const email = byId.get(entry.emailId);
+        if (email) {
+          // Conditional on the envelope still existing, so a body fetched just
+          // before its envelope was destroyed cannot land as an orphan.
+          const wrote = store.putBodyIfEnvelopeExists(jmapAccountId, entry.emailId, serialiseBody(email));
+          store.dequeueBodies(jmapAccountId, [entry.emailId]);
+          if (wrote) { fetched++; report.bodiesWritten++; }
+          continue;
+        }
+        if (notFound.has(entry.emailId)) {
+          // The message is gone, so the entry can NEVER succeed and must not burn
+          // five attempts. A durable give-up, not a row deletion - deleting it
+          // would let the backfill pass re-insert it every cycle forever.
+          store.markBodyGaveUp(jmapAccountId, [
+            { emailId: entry.emailId, receivedAt: entry.receivedAt, reason: 'notFound' },
+          ]);
+          continue;
+        }
+        // Present in neither list nor notFound: a transient miss.
+        const attempts = entry.attempts + 1;
+        if (attempts >= MAX_BODY_ATTEMPTS) {
+          store.markBodyGaveUp(jmapAccountId, [
+            {
+              emailId: entry.emailId,
+              receivedAt: entry.receivedAt,
+              reason: 'attempts',
+              lastError: `gave up after ${attempts} attempts`,
+            },
+          ]);
+        } else {
+          store.bumpBodyAttempt(
+            jmapAccountId, entry.emailId, Date.now() + backoffDelayMs(attempts),
+            'the server returned neither the record nor notFound',
+          );
+        }
+      }
+    });
+  }
+  return fetched;
+}
+
+// ── C2: body backfill ───────────────────────────────────────────────────────
+
+/**
+ * Notices envelopes that never had a body enqueued - a widened body window, or a
+ * queue row lost to a purge. Runs EVEN WHEN C1 found nothing, because that is its
+ * whole job.
+ *
+ * It is cap-aware and excludes durable give-ups. Both matter: the headroom check
+ * avoids paying for a download with nowhere to go, and the give-up exclusion is
+ * what makes termination provable. Without it, the size cap sheds the oldest
+ * bodies, those envelopes are still inside the body WINDOW, so this pass
+ * re-enqueues them, they download again, and the cap sheds them again - unbounded
+ * data use with no termination, because there is always an envelope without a body.
+ *
+ * A heuristic frontier instead of a durable mark does NOT work, and the reason is
+ * worth recording: the evictor sheds down TO the cap, which leaves headroom, so
+ * this pass refills and the two trade the same bytes back and forth. Only the
+ * durable mark makes the store monotone.
+ */
+async function backfillBodies(
+  ctx: CycleContext,
+  bodyFrom: string,
+  maxBodyBytes: number,
+  alreadyFetched: number,
+): Promise<number> {
+  const { store, jmapAccountId } = ctx;
+  const budget = BUDGET.bodyItems - alreadyFetched;
+  if (budget <= 0) return 0;
+
+  const used = store.bodyBytesTotal(jmapAccountId);
+  let headroom = maxBodyBytes - used;
+  if (headroom <= 0) return 0;
+
+  const candidates = store.envelopesWithoutBody(jmapAccountId, bodyFrom, budget * 2);
+  if (candidates.length === 0) return 0;
+  const gaveUp = new Set(store.listBodyGiveUps(jmapAccountId, 5_000));
+
+  const entries: BodyQueueEntry[] = [];
+  for (const c of candidates) {
+    if (entries.length >= budget) break;
+    if (gaveUp.has(c.id)) continue;
+    // Always allow at least one, or a single oversized message would stall the
+    // queue forever.
+    if (entries.length > 0 && c.size > headroom) break;
+    headroom -= c.size;
+    entries.push({ emailId: c.id, jmapAccountId, receivedAt: c.receivedAt, attempts: 0 });
+    if (headroom <= 0) break;
+  }
+  if (entries.length === 0) return 0;
+
+  // The INSERTED count, not the attempted count. Reporting the latter made the
+  // mobile engine treat every cycle as having unfinished work for as long as any
+  // envelope lacked a body, chaining a new cycle every few seconds indefinitely.
+  return store.transaction(() => store.enqueueBodies(entries));
+}
+
+// ── retention ────────────────────────────────────────────────────────────────
+
+function applyRetention(
+  ctx: CycleContext,
+  args: {
+    envelopeFrom: string;
+    bodyFrom: string;
+    maxBodyBytes: number;
+    evictionAllowed: boolean;
+    previousFloor: string | undefined;
+    report: CycleReport;
+  },
+): void {
+  const { store, jmapAccountId, now } = ctx;
+  const { envelopeFrom, bodyFrom, maxBodyBytes, evictionAllowed, previousFloor, report } = args;
+  const coverage = store.getCoverage(jmapAccountId);
+
+  // A retention change arriving DURING a reconcile is deferred to after the
+  // sweep. Applying a widen mid-reconcile would make the sweep delete everything
+  // between the old and new floors permanently; a narrow is deferred too, because
+  // evicting below a floor the sweep is about to use races it for no benefit.
+  if (coverage?.phase === 'reconciling') {
+    if (coverage.targetFrom !== envelopeFrom) {
+      store.transaction(() => {
+        store.patchCoverage(jmapAccountId, { deferredTargetFrom: envelopeFrom });
+      });
+    }
+    return;
+  }
+
+  const movement = floorMovement(previousFloor, envelopeFrom);
+  const adjustment = adjustForWindow(movement, envelopeFrom);
+
+  if (adjustment.evictBelow) {
+    if (!evictionAllowed) {
+      report.warnings.push(
+        `skipping retention eviction below ${adjustment.evictBelow}: the floor came from a ` +
+          `suppressed clock anomaly`,
+      );
+    } else {
+      store.transaction(() => {
+        report.envelopesDeleted += store.evictEnvelopesBelow(jmapAccountId, adjustment.evictBelow as string);
+        // `coveredFrom` follows the floor, but ONLY when it was already set:
+        // otherwise a narrow claims a range that was never enumerated, and delta
+        // sync cannot re-deliver pre-existing mail to repair that.
+        if (coverage?.coveredFrom !== null && coverage?.coveredFrom !== undefined) {
+          store.patchCoverage(jmapAccountId, { coveredFrom: adjustment.evictBelow as string });
+        }
+        store.patchCoverage(jmapAccountId, { targetFrom: envelopeFrom });
+      });
+    }
+  } else if (adjustment.rescanFrom) {
+    // A WIDEN is not a resync: the target moves back, coverage re-enters scanning,
+    // and the cursors are untouched.
+    store.transaction(() => {
+      store.patchCoverage(jmapAccountId, {
+        targetFrom: adjustment.rescanFrom as string,
+        scanCursor: adjustment.rescanFrom as string,
+        phase: 'scanning',
+      });
+    });
+    report.unfinishedWork = true;
+  }
+
+  // Body narrow: delete and merely DEQUEUE, never mark - the backfill's
+  // `receivedAfter` already excludes out-of-window bodies, and a later widen must
+  // be free to re-fetch them.
+  const stale = store.bodiesBelow(jmapAccountId, bodyFrom, 2_000);
+  if (stale.length > 0 && evictionAllowed) {
+    store.transaction(() => {
+      report.bodiesEvicted += store.deleteBodies(jmapAccountId, stale);
+      store.dequeueBodies(jmapAccountId, stale);
+    });
+  }
+
+  // Orphan bodies: invisible to cap eviction, which only walks the body table.
+  const orphans = store.orphanBodies(jmapAccountId, 500);
+  if (orphans.length > 0) {
+    store.transaction(() => {
+      report.bodiesEvicted += store.deleteBodies(jmapAccountId, orphans);
+      store.dequeueBodies(jmapAccountId, orphans);
+    });
+  }
+
+  // A cap RAISE revives bodies shed for space: a durable refusal recorded under
+  // one policy must not outlive that policy.
+  const flags = store.getFlags(now);
+  if (flags.lastMaxBodyBytes !== undefined && maxBodyBytes > flags.lastMaxBodyBytes) {
+    store.transaction(() => { store.clearBodyGiveUps(jmapAccountId, 'shed-by-cap'); });
+  }
+
+  // The MB cap: oldest bodies first, envelopes always survive, so the message
+  // stays listed and openable when back online.
+  let used = store.bodyBytesTotal(jmapAccountId);
+  if (used > maxBodyBytes) {
+    const oldest = store.oldestBodies(jmapAccountId, 5_000);
+    const shed: Array<{ emailId: string; receivedAt: string; reason: 'shed-by-cap' }> = [];
+    for (const b of oldest) {
+      if (used <= maxBodyBytes) break;
+      used -= b.bytes;
+      shed.push({ emailId: b.emailId, receivedAt: '', reason: 'shed-by-cap' });
+    }
+    if (shed.length > 0) {
+      // MARKED, not merely dequeued. See backfillBodies' comment for the loop
+      // this closes.
+      const withDates = shed.map((s) => {
+        const row = store.getEnvelopeRaw(jmapAccountId, s.emailId);
+        return { ...s, receivedAt: typeof row?.received_at === 'string' ? row.received_at : new Date(now).toISOString() };
+      });
+      store.transaction(() => {
+        report.bodiesEvicted += store.deleteBodies(jmapAccountId, shed.map((s) => s.emailId));
+        store.markBodyGaveUp(jmapAccountId, withDates);
+      });
+    }
+  }
+}
diff --git a/lib/offline-replica/types.ts b/lib/offline-replica/types.ts
new file mode 100644
index 00000000..884ffc4b
--- /dev/null
+++ b/lib/offline-replica/types.ts
@@ -0,0 +1,169 @@
+// Persisted shapes for the offline replica.
+//
+// Note that `SyncCursor.state` is declared as a plain `string` here while
+// `./states.ts` goes to some trouble to brand it. That is deliberate: a token
+// that has round-tripped through JSON has no provenance left to certify. The
+// brands guard the WRITE PATHS (`advanceCursor` / `seedCursor`), which is where
+// provenance is actually decided; the row is just a row.
+
+import type { CursorType } from './states';
+
+export interface SyncCursor {
+  type: CursorType;
+  jmapAccountId: string;
+  /** A ChangesState from this (type, jmapAccountId), or a seeded SnapshotState. */
+  state: string;
+  /** True when the last page reported `hasMoreChanges` - a drain is unfinished. */
+  drainPending: boolean;
+  /** Set when the server invalidated us. Cleared only by a COMPLETED reconcile. */
+  invalidatedAt?: number;
+  invalidatedReason?: 'cannotCalculateChanges' | 'oldStateMismatch' | 'corruptState' | 'manual';
+  /**
+   * Anti-wedge counters, PER CURSOR rather than per account. With the counters
+   * shared, a healthy Mailbox cursor resetting them every cycle meant a failing
+   * Email cursor never escalated and never advanced again - silently, forever.
+   */
+  consecutiveFailures: number;
+  /** The `sinceState` that failed; escalation only counts failures at the same position. */
+  lastFailedState?: string;
+  /** Current rung of the `maxChanges` ladder. */
+  maxChangesRung: 0 | 1 | 2 | 3;
+  updatedAt: number;
+}
+
+export interface CoverageState {
+  jmapAccountId: string;
+  /** ISO. Oldest `receivedAt` for which the ENVELOPE tier is known-complete. */
+  coveredFrom: string | null;
+  /** ISO. Ascending scan resume point; null when not scanning. */
+  scanCursor: string | null;
+  /** The retention floor this scan is working toward. */
+  targetFrom: string;
+  /** The floor PINNED at reconcile start. The sweep deletes only against this. */
+  sweepFloor?: string;
+  /** Set when a retention change arrived mid-reconcile; applied after the sweep. */
+  deferredTargetFrom?: string;
+  /**
+   * The `cached_at` stamp a reconcile's enumeration writes onto every envelope it
+   * re-sees, pinned when the reconcile starts.
+   *
+   * This is a multi-cycle, crash-resumable "seen set" implemented as ONE
+   * INTEGER, with no seen-ids table: the enumeration re-upserts each surviving
+   * envelope, refreshing its `cached_at`, so the sweep is
+   * `received_at >= sweepFloor AND cached_at < reconcileStampedAt`. A record the
+   * enumeration never reached keeps its older stamp and is swept; a record the
+   * LIVE delta path creates mid-reconcile gets `Date.now() >= stamp` and
+   * survives, which is exactly right.
+   *
+   * The stamp must be derived from the DATA, not the clock:
+   * `max(now, maxCachedAt + 1)`. With a frozen or coarse clock, `cached_at <
+   * stamp` matches nothing and the sweep deletes nothing.
+   */
+  reconcileStampedAt?: number;
+  /** Durable trace of any tie-cluster skip taken by the last-resort paging rung. */
+  gapMarkers?: Array<{ from: string; to: string; reason: 'tie-cluster-skip'; at: number }>;
+  phase: 'never-run' | 'scanning' | 'reconciling' | 'complete';
+  /** Progress, for the UI only. Never load-bearing. */
+  seen: number;
+  consecutiveFailures: number;
+  updatedAt: number;
+}
+
+export interface BodyQueueEntry {
+  emailId: string;
+  jmapAccountId: string;
+  /** Drives priority: newest first. */
+  receivedAt: string;
+  /** NEVER reset by a re-enqueue. */
+  attempts: number;
+  lastError?: string;
+  nextAttemptAt?: number;
+  /**
+   * Durable terminal state. A gave-up row is KEPT rather than deleted, precisely
+   * so the backfill job cannot resurrect it - its driver is "envelope without a
+   * body", which by itself cannot distinguish "not fetched yet" from
+   * "deliberately not kept". Cleared wholesale by a completed reconcile, so a
+   * transient outage self-heals.
+   */
+  gaveUp?: boolean;
+  gaveUpReason?: 'attempts' | 'notFound' | 'shed-by-cap';
+}
+
+export type BodyGiveUpReason = NonNullable<BodyQueueEntry['gaveUpReason']>;
+
+/** Per-account flags. A VIEW over field-level patches, never written whole. */
+export interface ReplicaFlags {
+  /** Sticky until a reconcile completes. Survives restarts. */
+  resyncRequired: boolean;
+  /** Rolling count + window start for the reconcile ceiling. */
+  reconcilesInWindow: number;
+  reconcileWindowStartedAt: number;
+  /**
+   * Last observed retention floor, for the clock-jump guard.
+   *
+   * This must hold the floor that was actually USED, never the suspicious one
+   * that was rejected - see `retention.ts` for the wipe that the other choice
+   * caused.
+   */
+  lastWindowFloor?: string;
+  /**
+   * The `envelopeDays` that produced `lastWindowFloor`.
+   *
+   * The computed floor moves for TWO independent reasons - the clock changing
+   * and the SETTING changing - and guarding a setting change is wrong: it is
+   * explicit user intent, not a glitch. Recording the policy alongside the floor
+   * is what tells them apart.
+   */
+  lastEnvelopeDays?: number;
+  /**
+   * The body-tier byte cap in force last cycle. A RAISE must revive bodies
+   * previously shed for space, which is otherwise a durable refusal.
+   */
+  lastMaxBodyBytes?: number;
+  lastCycleAt?: number;
+  lastCycleOk?: boolean;
+  lastCycleError?: string;
+}
+
+export function defaultFlags(now: number): ReplicaFlags {
+  return {
+    resyncRequired: false,
+    reconcilesInWindow: 0,
+    reconcileWindowStartedAt: now,
+  };
+}
+
+export type FlagsPatch = Partial<ReplicaFlags>;
+
+/** The envelope tier, as stored. */
+export interface EnvelopeRow {
+  jmapAccountId: string;
+  id: string;
+  threadId: string | null;
+  receivedAt: string;
+  size: number | null;
+  subject: string | null;
+  preview: string | null;
+  fromJson: string | null;
+  toJson: string | null;
+  ccJson: string | null;
+  blobId: string | null;
+  hasAttachment: boolean;
+  keywordsJson: string;
+  mailboxIds: string[];
+}
+
+export interface MailboxRow {
+  jmapAccountId: string;
+  id: string;
+  name: string;
+  parentId: string | null;
+  role: string | null;
+  sortOrder: number | null;
+  totalEmails: number | null;
+  unreadEmails: number | null;
+  totalThreads: number | null;
+  unreadThreads: number | null;
+  myRightsJson: string | null;
+  isSubscribed: boolean;
+}
diff --git a/playwright.integration-electron.config.ts b/playwright.integration-electron.config.ts
index 7a0a7b2f..557f9507 100644
--- a/playwright.integration-electron.config.ts
+++ b/playwright.integration-electron.config.ts
@@ -22,10 +22,14 @@ import { defineConfig } from '@playwright/test';
 export default defineConfig({
   testDir: './integration/tests',
   // 11 asserts the native notification bridge fires from a real push; 12
-  // asserts a real delivery reaches the encrypted local search index. 12 runs
-  // the REAL standalone-server boot (no ELECTRON_LOAD_URL), because that boot
-  // is what wires the index's store directory and its fd-3 key channel.
-  testMatch: /1[12]-electron-.*\.spec\.ts/,
+  // asserts a real delivery reaches the encrypted local search index; 13 asserts
+  // the offline mail REPLICA still serves a synced message's full body with the
+  // backend severed at the socket level. 12 and 13 run the REAL standalone-server
+  // boot (no ELECTRON_LOAD_URL), because that boot is what wires the store
+  // directory and the fd-3 key channel.
+  testMatch: /1[123]-electron-.*\.spec\.ts/,
+  // 13 syncs a real mailbox and then chains cycles, so it needs more than 90s;
+  // it sets its own per-test timeout, and this is the floor for the others.
   timeout: 90_000,
   expect: { timeout: 20_000 },
   fullyParallel: false,
diff --git a/playwright.integration.config.ts b/playwright.integration.config.ts
index 0aeb9979..baa19f5c 100644
--- a/playwright.integration.config.ts
+++ b/playwright.integration.config.ts
@@ -30,7 +30,11 @@ export default defineConfig({
   // Playwright image) has no Electron binary compatible with that
   // container's platform, so they must never be swept in by this config's
   // default testDir glob.
-  testIgnore: ['11-electron-notification.spec.ts', '12-electron-mail-index.spec.ts'],
+  testIgnore: [
+    '11-electron-notification.spec.ts',
+    '12-electron-mail-index.spec.ts',
+    '13-electron-offline-replica.spec.ts',
+  ],
   // next dev compiles routes lazily and each test logs in fresh, so give
   // individual tests and their polling assertions generous headroom.
   timeout: 90_000,
diff --git a/stores/auth-store.ts b/stores/auth-store.ts
index f302a0cc..6e11de84 100644
--- a/stores/auth-store.ts
+++ b/stores/auth-store.ts
@@ -1,6 +1,7 @@
 import { create } from 'zustand';
 import { persist } from 'zustand/middleware';
 import { JMAPClient, RateLimitError } from '@/lib/jmap/client';
+import { withOfflineFallback } from '@/lib/offline-fallback-client';
 import type { IJMAPClient } from '@/lib/jmap/client-interface';
 import { useIdentityStore } from './identity-store';
 import { setClientLookup } from './client-registry';
@@ -638,14 +639,14 @@ export const useAuthStore = create<AuthState>()(
             } else {
               // Legacy fallback for pre-0.16 Stalwart, which accepts the TOTP
               // appended to the password over basic auth.
-              client = new JMAPClient(serverUrl, username, `${password}$${totp}`);
+              client = withOfflineFallback(new JMAPClient(serverUrl, username, `${password}$${totp}`));
               await client.connect();
               const { useTotpReauthStore } = await import('@/stores/totp-reauth-store');
               client.enableTotpReauth(password, () => useTotpReauthStore.getState().requestTotp());
               debug.log('auth', 'TOTP re-auth enabled (legacy basic-auth path)');
             }
           } else {
-            client = new JMAPClient(serverUrl, username, password);
+            client = withOfflineFallback(new JMAPClient(serverUrl, username, password));
             await client.connect();
           }
 
@@ -1426,7 +1427,7 @@ export const useAuthStore = create<AuthState>()(
               const res = await apiFetch(`/api/auth/session?slot=${targetAccount.cookieSlot}`, { method: 'PUT' });
               if (res.ok) {
                 const { serverUrl, username, password } = await res.json();
-                targetClient = new JMAPClient(serverUrl, username, password);
+                targetClient = withOfflineFallback(new JMAPClient(serverUrl, username, password));
                 bindClientStatusHandlers(targetClient, set, get, accountId);
                 await targetClient.connect();
                 clients.set(accountId, targetClient);
@@ -1687,7 +1688,7 @@ export const useAuthStore = create<AuthState>()(
                 const res = await apiFetch(`/api/auth/session?slot=${account.cookieSlot}`, { method: 'PUT' });
                 if (res.ok) {
                   const { serverUrl, username, password } = await res.json();
-                  const client = new JMAPClient(serverUrl, username, password);
+                  const client = withOfflineFallback(new JMAPClient(serverUrl, username, password));
                   bindClientStatusHandlers(client, set, get, account.id);
                   await client.connect();
                   clients.set(account.id, client);
@@ -1897,7 +1898,7 @@ export const useAuthStore = create<AuthState>()(
                   throw new Error('Incomplete session data');
                 }
                 const { serverUrl, username, password } = data;
-                const client = new JMAPClient(serverUrl, username, password);
+                const client = withOfflineFallback(new JMAPClient(serverUrl, username, password));
                 await client.connect();
 
                 const accountId = generateAccountId(username, serverUrl);
diff --git a/stores/email-store.ts b/stores/email-store.ts
index f0b97c0e..3b0359bb 100644
--- a/stores/email-store.ts
+++ b/stores/email-store.ts
@@ -2847,6 +2847,22 @@ export const useEmailStore = create<EmailStore>((set, get) => ({
       // arrived, i.e. index everything except the delivery that triggered it.
       const scheduleIndexUpdate = () => {
         void (async () => {
+          // The offline REPLICA (lib/offline-replica/**) rides the same trigger.
+          // It is a SEPARATE subsystem from the search index: the index keeps a
+          // plain-text excerpt for retrieval, the replica keeps full bodies plus
+          // properly-provenanced /changes cursors so mail stays READABLE with no
+          // network. Both write the same encrypted file on separate connections,
+          // and both are request-scoped with no background worker. Unlike the
+          // index, the replica DOES care about Mailbox changes - it holds the
+          // folder counters.
+          try {
+            const { syncOnStateChange } = await import('@/lib/offline-replica-client');
+            syncOnStateChange(change, {
+              slot: useAccountStore.getState().getActiveAccount()?.cookieSlot,
+            });
+          } catch {
+            /* the replica is optional; never let it affect mail handling */
+          }
           try {
             const { indexOnStateChange } = await import('@/lib/mail-index-client');
             const mailIds = get().emails.slice(0, 100).map((e) => e.id);