fix(notifications): clear thread receipts on mark-read; cap avatar-decoration refetch

Two federated-room bugs surfaced by the desktop build:

1. markAsRead only sent one unthreaded receipt at the main-timeline tail. With
   threadSupport enabled, thread replies leave the main timeline, so a reply
   newer than that tail was never covered — its per-thread notification count
   (which the room dot sums) lingered, so the unread dot never cleared even
   after reading. It also early-returned when the main timeline was already
   read. Now also send a threaded receipt at each unread thread's latest reply.

2. useAvatarDecoration never cached non-404 failures, so every avatar mount
   re-requested io.lotus.avatar_decoration for federated users whose homeserver
   403s/502s the field — a refetch storm that spammed the console and hammered
   our homeserver's federation. Now cache definitive rejections (400/403/404)
   and give up after ~2 transient (429/5xx) attempts per session.

Gates: tsc/eslint/prettier clean, build OK, 665 tests.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/app/hooks/useAvatarDecoration.ts

@@ -9,6 +9,9 @@ const PROFILE_FIELD = 'io.lotus.avatar_decoration';
 const cache = new Map<string, string | null>();
 // Callbacks waiting for a userId's result
 const pending = new Map<string, Array<(val: string | null) => void>>();
+// Transient-failure attempt counts (userId → n) so a flaky federated lookup
+// can retry a couple of times, then gives up for the session.
+const failures = new Map<string, number>();
 
 function fetchDecoration(
   authedRequest: (method: Method, path: string) => Promise<Record<string, string>>,
@@ -33,16 +36,23 @@ function fetchDecoration(
       return val;
     })
     .catch((err: unknown) => {
-      // A 404 (M_NOT_FOUND) means the field is genuinely unset → cache "no
-      // decoration". A transient failure (429 rate-limit, 5xx, network) must
-      // NOT be cached: doing so permanently hides the user's decoration for the
-      // whole session. This matters most for the member list and timeline, which
-      // mount many avatars at once and can trip homeserver rate limits — a
-      // single 429 in that burst would otherwise wipe the decoration until a
-      // full reload. Leaving the cache unset lets the next mount retry.
       const status = err instanceof MatrixError ? err.httpStatus : undefined;
-      if (status === 404) {
+      // Definitive rejections — the field is unset (404) or the server won't
+      // serve it (400/403). This is the common case for FEDERATED users whose
+      // homeserver doesn't support extended profiles / rejects the field. Cache
+      // "no decoration" so we never refetch: otherwise every avatar mount
+      // re-requests and floods our homeserver with failing federated profile
+      // lookups (the 403/502 console storm + real HS load).
+      if (status === 404 || status === 403 || status === 400) {
         cache.set(userId, null);
+      } else {
+        // Transient (429 rate-limit / 5xx / network). Allow a couple of retries
+        // — a single 429 in a member-list burst shouldn't permanently hide a
+        // decoration — then give up for the session so a persistently-failing
+        // federated link (e.g. a 502'ing remote server) can't loop forever.
+        const attempts = (failures.get(userId) ?? 0) + 1;
+        failures.set(userId, attempts);
+        if (attempts >= 2) cache.set(userId, null);
       }
       return null;
     })