diff --git a/LOTUS_TESTING.md b/LOTUS_TESTING.md
index c32b806d8..be069fd7c 100644
--- a/LOTUS_TESTING.md
+++ b/LOTUS_TESTING.md
@@ -677,8 +677,9 @@ Run the axe DevTools extension (or Lighthouse → Accessibility) on a room view,
 
 **Unread dot on federated rooms + avatar-decoration console storm (2026-07):**
 
-- Open a room from another homeserver that has thread activity; read it → the room's unread **dot clears** (previously an unread _thread reply_ kept the dot because `markAsRead` only sent an unthreaded receipt at the main-timeline tail). Also confirm opening a thread + reading it clears its part of the badge.
-- With DevTools console open on those rooms, the `io.lotus.avatar_decoration` `403`/`502` (and federated media) errors should **not** repeat on every scroll/mount — each failing user is now requested at most ~twice per session, so the storm (and its homeserver load) is gone.
+- **Read receipts (regression guard — highest priority):** open several rooms and open the Home/Direct tabs (which mark all orphan rooms read on mount) → rooms **stay read**, unread dots clear and don't come back. (A prior attempt sent a receipt for the thread _root_ when a thread's replies weren't loaded, which the SDK treats as a main receipt at an old event and re-unread every room on every mark-read. Fixed + locked by `notifications.test.ts`.)
+- **Thread dot:** a room with an unread reply in a thread whose replies are loaded → its dot clears on read; for a thread not yet loaded, the dot clears once you open/load the thread. (mark-as-read now sends a threaded receipt only for a genuine loaded reply, never the root.)
+- With DevTools console open on federated rooms, the `io.lotus.avatar_decoration` `403`/`502` (and federated media) errors should **not** repeat on every scroll/mount — each failing user is now requested at most ~twice per session, so the storm (and its homeserver load) is gone.
 
 **Custom Window Chrome (Beta) fix (2026-07):** on the desktop build, Settings → General → toggle **Custom Window Chrome** — it should reload and come up with the Lotus title bar and a normal, stable feed (no screen-expand / auto-scroll-into-the-past). Toggle back off → reloads to the native frame.
 
diff --git a/src/app/utils/notifications.test.ts b/src/app/utils/notifications.test.ts
new file mode 100644
index 000000000..2a952463f
--- /dev/null
+++ b/src/app/utils/notifications.test.ts
@@ -0,0 +1,126 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { NotificationCountType, ReceiptType } from 'matrix-js-sdk';
+import { markAsRead } from './notifications';
+
+// markAsRead sends an unthreaded read receipt at the latest main-timeline event,
+// plus a THREADED receipt at each unread thread's latest loaded reply. The
+// regression these tests guard against: a thread whose replies aren't loaded
+// (lastReply() === null) must NOT produce a receipt for the thread root — that
+// resolves to a MAIN receipt at an old event and permanently unreads the room.
+
+type ReceiptCall = { eventId: string; receiptType: ReceiptType; unthreaded?: boolean };
+
+const evt = (id: string, sending = false) => ({ getId: () => id, isSending: () => sending }) as any;
+
+const thread = (id: string, lastReply: any) => ({ id, lastReply: () => lastReply }) as any;
+
+type RoomOpts = {
+  timeline?: any[];
+  readUpTo?: string | null;
+  threads?: any[];
+  threadUnread?: Record<string, number>;
+};
+
+const setup = (opts: RoomOpts) => {
+  const calls: ReceiptCall[] = [];
+  const room = {
+    getLiveTimeline: () => ({ getEvents: () => opts.timeline ?? [] }),
+    getEventReadUpTo: () => opts.readUpTo ?? null,
+    getThreads: () => opts.threads ?? [],
+    getThreadUnreadNotificationCount: (threadId: string, _type: NotificationCountType) =>
+      opts.threadUnread?.[threadId] ?? 0,
+  };
+  const mx = {
+    getRoom: () => room,
+    getUserId: () => '@me:server',
+    sendReadReceipt: async (event: any, receiptType: ReceiptType, unthreaded?: boolean) => {
+      calls.push({ eventId: event.getId(), receiptType, unthreaded });
+      return {};
+    },
+  } as any;
+  return { mx, calls };
+};
+
+test('main timeline: unthreaded receipt at the latest event', async () => {
+  const { mx, calls } = setup({ timeline: [evt('a'), evt('b'), evt('c')], readUpTo: 'a' });
+  await markAsRead(mx, '!r:server', false);
+  assert.equal(calls.length, 1);
+  assert.deepEqual(calls[0], { eventId: 'c', receiptType: ReceiptType.Read, unthreaded: true });
+});
+
+test('REGRESSION: an unread thread with unloaded replies (lastReply null) sends NO root receipt', async () => {
+  const t = thread('$root', null); // replies not loaded
+  const { mx, calls } = setup({
+    timeline: [evt('a'), evt('b')],
+    readUpTo: 'a',
+    threads: [t],
+    threadUnread: { $root: 3 },
+  });
+  await markAsRead(mx, '!r:server', false);
+  // Only the main unthreaded receipt — never a receipt for the thread root.
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].eventId, 'b');
+  assert.equal(calls[0].unthreaded, true);
+  assert.ok(!calls.some((c) => c.eventId === '$root'));
+});
+
+test('unread thread with a loaded reply sends a threaded receipt at that reply', async () => {
+  const t = thread('$root', evt('$reply'));
+  const { mx, calls } = setup({
+    timeline: [evt('a'), evt('b')],
+    readUpTo: 'a',
+    threads: [t],
+    threadUnread: { $root: 1 },
+  });
+  await markAsRead(mx, '!r:server', false);
+  const main = calls.find((c) => c.eventId === 'b');
+  const threaded = calls.find((c) => c.eventId === '$reply');
+  assert.ok(main && main.unthreaded === true);
+  assert.ok(threaded && threaded.unthreaded === false);
+  assert.equal(calls.length, 2);
+});
+
+test('main already read but a thread is unread: no main receipt, threaded receipt only', async () => {
+  const t = thread('$root', evt('$reply'));
+  const { mx, calls } = setup({
+    timeline: [evt('a'), evt('b')],
+    readUpTo: 'b', // latest main event already read → getLatestValidEvent() null
+    threads: [t],
+    threadUnread: { $root: 2 },
+  });
+  await markAsRead(mx, '!r:server', false);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].eventId, '$reply');
+  assert.equal(calls[0].unthreaded, false);
+});
+
+test('everything read: no receipts sent', async () => {
+  const t = thread('$root', evt('$reply'));
+  const { mx, calls } = setup({
+    timeline: [evt('a'), evt('b')],
+    readUpTo: 'b',
+    threads: [t],
+    threadUnread: { $root: 0 }, // thread read too
+  });
+  await markAsRead(mx, '!r:server', false);
+  assert.equal(calls.length, 0);
+});
+
+test('sending thread reply is skipped', async () => {
+  const t = thread('$root', evt('$reply', true)); // isSending → skip
+  const { mx, calls } = setup({
+    timeline: [evt('a'), evt('b')],
+    readUpTo: 'b',
+    threads: [t],
+    threadUnread: { $root: 1 },
+  });
+  await markAsRead(mx, '!r:server', false);
+  assert.equal(calls.length, 0);
+});
+
+test('private receipt flag uses ReadPrivate', async () => {
+  const { mx, calls } = setup({ timeline: [evt('a'), evt('b')], readUpTo: 'a' });
+  await markAsRead(mx, '!r:server', true);
+  assert.equal(calls[0].receiptType, ReceiptType.ReadPrivate);
+});
diff --git a/src/app/utils/notifications.ts b/src/app/utils/notifications.ts
index 2dbaca6bb..2156d3415 100644
--- a/src/app/utils/notifications.ts
+++ b/src/app/utils/notifications.ts
@@ -25,25 +25,33 @@ export async function markAsRead(mx: MatrixClient, roomId: string, privateReceip
   if (latestEvent) {
     // Unthreaded receipt: with client threadSupport enabled the SDK would
     // otherwise scope this to the main timeline (thread_id: "main"). Unthreaded
-    // clears the main timeline + every event up to this one, in any thread.
+    // clears the main timeline + every event up to this one.
     await mx.sendReadReceipt(latestEvent, receiptType, true);
   }
 
-  // ...but a thread reply NEWER than the main-timeline tail is not covered by
-  // the receipt above (threadSupport moves thread replies out of the main
-  // timeline), so its per-thread notification count — and the room's unread dot,
-  // which sums thread counts — would linger even after "reading" the room. Send
-  // a threaded receipt at the latest reply of every thread that still has unread
-  // counts. Also runs when the main timeline is already read (latestEvent null).
+  // Clear per-thread notification counts too — the room's unread dot sums them,
+  // so an unread thread reply keeps the dot lit even after the main timeline is
+  // read (threadSupport moves thread replies out of the main timeline, so the
+  // unthreaded receipt above doesn't necessarily cover them).
+  //
+  // CRITICAL: only send for a GENUINE loaded thread reply, via thread.lastReply().
+  // NEVER fall back to the thread root: a root event is "in the main timeline",
+  // so sendReadReceipt(root, false) resolves (via threadIdForReceipt) to a MAIN
+  // receipt at that old root event. If the root isn't in the loaded timeline it
+  // moves the main read receipt onto an event we don't have -> getEventReadUpTo()
+  // returns null -> the room is reported unread on every mark-read call (this was
+  // the P6 regression, amplified by the bulk mark-all-orphan-rooms-read callers).
+  // If a thread's replies aren't loaded (lastReply() null), just skip it.
   const threads = room.getThreads();
   await Promise.all(
     threads.map((thread) => {
       const unread =
         room.getThreadUnreadNotificationCount(thread.id, NotificationCountType.Total) ?? 0;
       if (unread <= 0) return undefined;
-      const lastReply = thread.lastReply() ?? thread.rootEvent;
+      const lastReply = thread.lastReply();
       if (!lastReply || lastReply.isSending()) return undefined;
-      // Threaded receipt (unthreaded = false → the SDK scopes it to this thread).
+      // Threaded receipt (unthreaded = false → the SDK scopes it to this thread
+      // via the reply's real threadRootId; it never touches the main marker).
       return mx.sendReadReceipt(lastReply, receiptType, false).catch(() => undefined);
     }),
   );