fix(discover): let a refreshed row override a stale local watch patch

Pausing an episode on one device, finishing it on another and pressing
Refresh left the first device showing the old "minutes left". Restarting the
app showed the right value. Two independent defects produce that, and either
alone reproduces the report.

The first is the watch-state overlay. Every local watch event lands in
WatchStateStore as a patch, and WatchStateSnapshot.apply overwrites
viewOffsetMs unconditionally; isNewerThan only ever orders one patch against
another, never against the server row underneath. Nothing expires a patch and
nothing clears the map except a profile switch, so the Mac's own paused
position kept winning over every subsequent fetch until the process died.

A patch exists to bridge the gap between a local action and the next server
read of that item, so it should stop applying once that read happens. The
store now records the watermark at which a successful authoritative response
returned each key, and suppresses an acknowledged session patch at or below
it. Only a watermark is stored, never the observed state: WatchStateSnapshot
cannot hold a container's leaf counts, and keeping max() per key makes the
order two concurrent responses complete irrelevant. Suppression is a
read-time predicate, so nothing mutates during build.

The barrier covers the parentChain too. patchForItem picks the newest of the
item's own entry and its ancestors', so retiring only the item's entry would
let an older season mark win and render watched/0 -- worse than either the
stale value or the fresh one. An authoritative read of a child already
reflects any container mark that preceded it, so the child's observation
judges its ancestors as well; a newer container action still wins.

Provenance decides what may be suppressed at all. WatchStateEvent now carries
serverAcknowledged, defaulting to false so an unclassified emit site degrades
to today's behaviour rather than silently becoming retireable. An offline
write is owed to the server and a read must never retire it, so it stays
until a WatchPatchPromotionNotifier promotion says the queue replayed it. That
channel is deliberately not a WatchStateEvent: OfflineWatchSyncService reacts
to watched/unwatched by purging queued progress, so replaying one there would
delete a newer rewatch. Promotion matches an exact WatchPatchId -- session
minted for live crossings, derived from the persisted (profile, row, revision)
for queued ones so it still joins after a restart.

Report acceptance is not delivery: PlaybackReportSession resolves true for a
same-state startup heartbeat it drops, so acknowledgement now keys on
onDelivered. A MediaBrowser Started saves play count and last-played date but
not the position, so it cannot acknowledge an offset. No report-derived
watched crossing is acknowledged on any backend -- Jellyfin hard-codes its
threshold and Plex never loads the server pref that would tell it the real
one -- so only an awaited explicit markWatched settles one.

The second defect is that a failed Refresh reported success. Plex _fetchHubs
and the Jellyfin hub legs both degrade a failure to an empty list, and the
library prefetch discarded its failures, so a server whose every hub request
failed was recorded as succeeded; DiscoverProvider then kept the previous rows,
set loaded and surfaced nothing. Worse, the background Continue Watching
refresh wiped the row outright on zero success.

Hub legs now report what they degraded through a HubFetchDiagnostics sink,
which keeps partial rows alongside the failure and leaves every existing
caller untouched. Failures ride through the aggregation results, a leg that
could not run because discovery failed contributes that failure rather than a
successful no-op, and loaded-server ids became succeeded - failed - cancelled
so one bad leg no longer caches a server as covered and blocks its retry. The
toolbar awaits a DiscoverRefreshOutcome and shows the existing unableToLoad
snackbar on failure while the retained rows stay on screen. Rollback after a
mid-pass exception is version-guarded, refilters against the current hidden
libraries and no longer publishes a system shelf the pass never committed.

Observations are staged with the pass and flushed only once the same disposal,
generation and exception checks that authorise committing those rows have
passed, so a discarded or rolled-back response can never suppress a patch.

Also fixes a live data-loss race the promotion work would have built on:
upsertProgressAction stamped a millisecond timestamp and updated the row in
place, so a rewatch queued during an in-flight replay was deleted by id.
Revisions are now strictly monotonic per row, replay deletes and retry updates
compare against them, and the upsert resets the retry fields because a new
revision is a new logical action.

close #1829

lib/database/app_database.dart

@@ -882,7 +882,7 @@ class AppDatabase extends _$AppDatabase {
   }
 
   /// Insert or update a progress action (merges with existing).
-  Future<void> upsertProgressAction({
+  Future<({int rowId, int revision})> upsertProgressAction({
     String? profileId,
     required ServerId serverId,
     String? clientScopeId,
@@ -895,7 +895,7 @@ class AppDatabase extends _$AppDatabase {
       final globalKey = buildGlobalKey(ServerId(serverId), ratingKey);
       final now = DateTime.now().millisecondsSinceEpoch;
 
-      await transaction(() async {
+      return transaction(() async {
         final existing =
             await (select(offlineWatchProgress)
                   ..where(
@@ -910,6 +910,12 @@ class AppDatabase extends _$AppDatabase {
 
         final keep = existing.isEmpty ? null : existing.first;
         if (keep != null) {
+          // The row id survives a merge, so its timestamp must advance even
+          // when multiple playback updates land in one clock millisecond.
+          final nextRevision = keep.updatedAt + 1;
+          final revision = now > nextRevision ? now : nextRevision;
+          // A merge is a new logical action; retry history belongs only to
+          // the revision whose server write failed.
           await (update(offlineWatchProgress)..where((t) => t.id.equals(keep.id))).write(
             OfflineWatchProgressCompanion(
               viewOffset: Value(viewOffset),
@@ -917,15 +923,19 @@ class AppDatabase extends _$AppDatabase {
               shouldMarkWatched: Value(shouldMarkWatched),
               profileId: Value(profileId),
               clientScopeId: Value(clientScopeId),
-              updatedAt: Value(now),
+              updatedAt: Value(revision),
+              syncAttempts: const Value(0),
+              lastError: const Value<String?>(null),
             ),
           );
           final duplicateIds = existing.skip(1).map((row) => row.id).toList(growable: false);
           if (duplicateIds.isNotEmpty) {
             await (delete(offlineWatchProgress)..where((t) => t.id.isIn(duplicateIds))).go();
           }
-        } else {
+          return (rowId: keep.id, revision: revision);
-          await into(offlineWatchProgress).insert(
+        }
+
+        final rowId = await into(offlineWatchProgress).insert(
           OfflineWatchProgressCompanion.insert(
             serverId: serverId,
             profileId: Value(profileId),
@@ -940,14 +950,14 @@ class AppDatabase extends _$AppDatabase {
             updatedAt: now,
           ),
         );
-        }
+        return (rowId: rowId, revision: now);
       });
     });
   }
 
   /// Insert a manual watch action (watched or unwatched).
   /// Removes conflicting actions for the same item.
-  Future<void> insertWatchAction({
+  Future<({int rowId, int revision})> insertWatchAction({
     String? profileId,
     required ServerId serverId,
     String? clientScopeId,
@@ -958,7 +968,7 @@ class AppDatabase extends _$AppDatabase {
       final globalKey = buildGlobalKey(ServerId(serverId), ratingKey);
       final now = DateTime.now().millisecondsSinceEpoch;
 
-      await transaction(() async {
+      return transaction(() async {
         // Remove conflicting actions (opposite action type and progress).
         await (delete(offlineWatchProgress)..where(
               (t) =>
@@ -968,7 +978,7 @@ class AppDatabase extends _$AppDatabase {
             ))
             .go();
 
-        await into(offlineWatchProgress).insert(
+        final rowId = await into(offlineWatchProgress).insert(
           OfflineWatchProgressCompanion.insert(
             serverId: serverId,
             profileId: Value(profileId),
@@ -980,6 +990,7 @@ class AppDatabase extends _$AppDatabase {
             updatedAt: now,
           ),
         );
+        return (rowId: rowId, revision: now);
       });
     });
   }
@@ -994,13 +1005,15 @@ class AppDatabase extends _$AppDatabase {
   /// orders by `createdAt` — and replaying it rewrites the resume position the
   /// mark just cleared, pinning the item to Continue Watching (#1812).
   ///
-  /// Progress queued *after* a mark is a genuine rewatch and is not affected:
+  /// When [beforeRevision] is present, the notifier is settling a persisted
-  /// this only runs at the moment the mark lands.
+  /// offline mark. Its listener is asynchronous, so only older revisions are
+  /// stale; an equal or newer progress revision is a genuine concurrent rewatch.
   Future<int> deleteQueuedProgressForItem({
     String? profileId,
     required ServerId serverId,
     String? clientScopeId,
     required String ratingKey,
+    int? beforeRevision,
   }) {
     return _runPendingMutation(() async {
       final globalKey = buildGlobalKey(ServerId(serverId), ratingKey);
@@ -1009,29 +1022,58 @@ class AppDatabase extends _$AppDatabase {
                 t.globalKey.equals(globalKey) &
                 _nullableTextPredicate(t.profileId, profileId) &
                 _nullableTextPredicate(t.clientScopeId, clientScopeId) &
-                t.actionType.equals(OfflineActionType.progress.id),
+                t.actionType.equals(OfflineActionType.progress.id) &
+                (beforeRevision == null ? const Constant(true) : t.updatedAt.isSmallerThanValue(beforeRevision)),
           ))
           .go();
     });
   }
 
-  /// Delete a specific watch action after successful sync
+  /// Delete a watch action only if it is still the snapshotted revision.
+  Future<bool> deleteWatchActionIfUnchanged(int id, int revision) {
+    return _runPendingMutation(() async {
+      final deleted = await (delete(
+        offlineWatchProgress,
+      )..where((t) => t.id.equals(id) & t.updatedAt.equals(revision))).go();
+      return deleted != 0;
+    });
+  }
+
+  /// Update the retry state only if the action is still the snapshotted revision.
+  Future<bool> updateSyncAttemptIfUnchanged(int id, int revision, String? errorMessage) {
+    return _runPendingMutation(() async {
+      final existing = await (select(
+        offlineWatchProgress,
+      )..where((t) => t.id.equals(id) & t.updatedAt.equals(revision))).getSingleOrNull();
+      if (existing == null) return false;
+
+      final updated = await (update(offlineWatchProgress)..where((t) => t.id.equals(id) & t.updatedAt.equals(revision)))
+          .write(
+            OfflineWatchProgressCompanion(
+              syncAttempts: Value(existing.syncAttempts + 1),
+              lastError: Value(errorMessage),
+            ),
+          );
+      return updated != 0;
+    });
+  }
+
+  /// Delete a specific watch action outside a snapshotted replay.
   Future<void> deleteWatchAction(int id) {
     return _runPendingMutation(() async {
       await (delete(offlineWatchProgress)..where((t) => t.id.equals(id))).go();
     });
   }
 
-  /// Update sync attempt count and error message
+  /// Update retry state outside a snapshotted replay.
-  Future<void> updateSyncAttempt(int id, String? errorMessage) async {
+  Future<void> updateSyncAttempt(int id, String? errorMessage) {
     return _runPendingMutation(() async {
       final existing = await (select(offlineWatchProgress)..where((t) => t.id.equals(id))).getSingleOrNull();
+      if (existing == null) return;
 
-      if (existing != null) {
       await (update(offlineWatchProgress)..where((t) => t.id.equals(id))).write(
         OfflineWatchProgressCompanion(syncAttempts: Value(existing.syncAttempts + 1), lastError: Value(errorMessage)),
       );
-      }
     });
   }
 

lib/media/media_server_client.dart

@@ -70,6 +70,42 @@ abstract interface class GracefullyCloseable {
   Future<void> closeGracefully({Duration drainTimeout});
 }
 
+/// Per-leg outcome sink for hub fetches.
+///
+/// Home rows are best-effort by design: a hub leg that fails degrades to an
+/// empty list rather than sinking the screen. That makes "this server has no
+/// rows" and "every row request failed" indistinguishable at the aggregation
+/// boundary, so a totally failed refresh was reported to the user as a
+/// success (#1829).
+///
+/// Callers that need to tell them apart pass a sink and the backend records
+/// each leg it degraded. A sink rather than a widened return type keeps every
+/// existing caller untouched, and lets a response carry *partial* rows
+/// alongside a failure — which throwing cannot.
+///
+/// Deliberately not recorded: legs whose degradation is intentional rather
+/// than a fault, i.e. Emby's series-reconstruction deadline and its
+/// per-series probes.
+class HubFetchDiagnostics {
+  var _failed = false;
+  var _cancelled = false;
+
+  /// A leg failed for a reason that is not a client-side abort.
+  bool get failed => _failed;
+
+  /// A leg was aborted client-side. Cancellation is not failure: a disrupted
+  /// pass says nothing about the server's actual content.
+  bool get cancelled => _cancelled;
+
+  void recordFailure(Object error) {
+    if (error is MediaServerHttpException && error.isCancellation) {
+      _cancelled = true;
+    } else {
+      _failed = true;
+    }
+  }
+}
+
 abstract class MediaServerClient {
   ServerId get serverId;
   String? get serverName;
@@ -302,7 +338,14 @@ abstract class MediaServerClient {
 
   /// Curated home-screen hubs across all libraries (Plex Discover; Jellyfin
   /// synthesizes `Latest` plus optional `Resume` + `NextUp`).
-  Future<List<MediaHub>> fetchGlobalHubs({int limit = defaultHubPreviewLimit, bool includePlaybackHubs = true});
+  ///
+  /// [diagnostics], when supplied, receives every leg this call degraded to
+  /// empty, so the caller can tell an empty home from a failed one (#1829).
+  Future<List<MediaHub>> fetchGlobalHubs({
+    int limit = defaultHubPreviewLimit,
+    bool includePlaybackHubs = true,
+    HubFetchDiagnostics? diagnostics,
+  });
 
   /// Hubs scoped to a single library section. [libraryName] is baked into
   /// the title of synthetic hubs (Jellyfin) so per-library "Recently Added"
@@ -310,12 +353,14 @@ abstract class MediaServerClient {
   /// [includePlaybackHubs] lets surfaces that already render Continue
   /// Watching skip duplicate playback rows. [libraryKind] lets backends avoid
   /// irrelevant expensive probes, e.g. Jellyfin `NextUp` for movie libraries.
+  /// [diagnostics] carries degraded legs as in [fetchGlobalHubs].
   Future<List<MediaHub>> fetchLibraryHubs(
     String libraryId, {
     required String libraryName,
     int limit = defaultHubPreviewLimit,
     bool includePlaybackHubs = true,
     MediaKind? libraryKind,
+    HubFetchDiagnostics? diagnostics,
   });
 
   /// "More like this" recommendations for [id].
@@ -769,10 +814,18 @@ extension MediaServerClientScope on MediaServerClient {
   /// In-player sessions that *did* give the backend an observable crossing use
   /// [notifyWatchedFromPlaybackSession] instead.
   Future<void> markWatchedFromPlaybackStop(MediaItem item) async {
-    if (!marksWatchedOnPlaybackStopped) {
+    final performedExplicitMark = !marksWatchedOnPlaybackStopped;
+    if (performedExplicitMark) {
       await markWatched(item);
     }
-    WatchStateNotifier().notifyWatched(item: item, isNowWatched: true, cacheServerId: cacheServerId);
+    WatchStateNotifier().notifyWatched(
+      item: item,
+      isNowWatched: true,
+      cacheServerId: cacheServerId,
+      // A successful report cannot prove that the backend marked the item
+      // played; only the explicit mutation settles this patch.
+      serverAcknowledged: performedExplicitMark,
+    );
   }
 
   /// Emit the local watched event for [item] without touching the server.
@@ -786,8 +839,17 @@ extension MediaServerClientScope on MediaServerClient {
   /// records the same watch twice — a second Trakt-plugin scrobble on Jellyfin
   /// (#1287), a second Play History row and an inflated `viewCount` on Plex
   /// (#1740).
-  void notifyWatchedFromPlaybackSession(MediaItem item) {
+  ///
-    WatchStateNotifier().notifyWatched(item: item, isNowWatched: true, cacheServerId: cacheServerId);
+  /// The event remains unacknowledged because reporting success proves only
+  /// receipt, not that the server classified the item as played. The caller
+  /// keeps the returned patch id so a later explicit mark can promote it.
+  WatchPatchId? notifyWatchedFromPlaybackSession(MediaItem item) {
+    return WatchStateNotifier().notifyWatched(
+      item: item,
+      isNowWatched: true,
+      cacheServerId: cacheServerId,
+      serverAcknowledged: false,
+    );
   }
 }
 

lib/navigation/profile_session_screen.dart

@@ -225,6 +225,10 @@ class _ProfileSessionScreenState extends State<ProfileSessionScreen> {
                     context.read<MultiServerProvider>(),
                     context.read<HiddenLibrariesProvider>(),
                     context.read<LibrariesProvider>(),
+                    // Created above in this same subtree, so its lifetime
+                    // matches; the proxy reuses `previous`, so the reference
+                    // stays valid for as long as this provider does.
+                    watchStateStore: context.read<WatchStateStore>(),
                     isProfileBinding: () => activeProfile.isBinding,
                     profileId: activeId,
                   );

lib/providers/discover_provider.dart

@@ -15,14 +15,46 @@ import '../utils/app_logger.dart';
 import '../utils/coalesced_load_coordinator.dart';
 import '../utils/deletion_notifier.dart';
 import '../utils/media_event_keys.dart';
+import '../utils/global_key_utils.dart';
 import '../utils/media_hub_ordering.dart';
 import '../utils/watch_state_notifier.dart';
 import 'hidden_libraries_provider.dart';
 import 'libraries_provider.dart';
 import 'multi_server_provider.dart';
+import 'watch_state_store.dart';
 
 enum DiscoverLoadState { initial, loading, loaded, error }
 
+enum DiscoverRefreshOutcome {
+  /// Both surfaces completed without a failed or cancelled server leg.
+  refreshed,
+
+  /// Some content refreshed, but at least one server leg failed or was cancelled.
+  degraded,
+
+  /// At least one server leg failed and none succeeded.
+  failed,
+
+  /// The pass was interrupted or had no attempted server legs.
+  cancelled,
+}
+
+DiscoverRefreshOutcome _refreshOutcome({
+  required Set<String> succeededServerIds,
+  required Set<String> failedServerIds,
+  required Set<String> cancelledServerIds,
+  required bool cancelled,
+}) {
+  if (cancelled) return DiscoverRefreshOutcome.cancelled;
+  if (failedServerIds.isNotEmpty) {
+    return succeededServerIds.isEmpty ? DiscoverRefreshOutcome.failed : DiscoverRefreshOutcome.degraded;
+  }
+  if (cancelledServerIds.isNotEmpty) {
+    return succeededServerIds.isEmpty ? DiscoverRefreshOutcome.cancelled : DiscoverRefreshOutcome.degraded;
+  }
+  return succeededServerIds.isEmpty ? DiscoverRefreshOutcome.cancelled : DiscoverRefreshOutcome.refreshed;
+}
+
 /// Owns the Discover tab's data: the Continue Watching row and the home hub
 /// list, including the refresh policy that used to live in the screen —
 /// durable watch events refresh only Continue Watching (one on-deck call,
@@ -48,8 +80,12 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
     this._libraries, {
     required this.profileId,
     required this.isProfileBinding,
+    WatchStateStore? watchStateStore,
     Future<void> Function(String profileId, List<MediaItem>)? syncSystemShelf,
-  }) : _syncSystemShelfOverride = syncSystemShelf {
+    // A private field cannot be a named initializing formal callers can pass.
+    // ignore: prefer_initializing_formals
+  }) : _watchStateStore = watchStateStore,
+       _syncSystemShelfOverride = syncSystemShelf {
     _loadCoordinator = CoalescedLoadCoordinator<String>(onFull: _loadOnce, onDelta: _loadDeltaOnce);
     // Late server connects (reconnect after outage, slow wave) refresh
     // discover the same way they refresh libraries. Removed in [dispose] so a
@@ -79,8 +115,46 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
   final MultiServerProvider _multiServer;
   final HiddenLibrariesProvider _hiddenLibraries;
   final LibrariesProvider _libraries;
+
+  /// Authoritative fetches tell the store which items the server re-observed,
+  /// so a stale local watch patch stops overriding fresh server state (#1829).
+  /// Optional: tests and isolated subtrees may have no store.
+  final WatchStateStore? _watchStateStore;
   final String? profileId;
 
+  /// Watermark plus store identity captured before an authoritative request.
+  /// Null when there is no store to reconcile against.
+  ({int watermark, Object epoch})? _beginObservation() {
+    final store = _watchStateStore;
+    if (store == null) return null;
+    return (watermark: store.observationWatermark, epoch: store.observationEpoch);
+  }
+
+  /// Tell the store which items a *successful and committed* pass re-observed,
+  /// so their stale local watch patches stop overriding the fresh server rows.
+  ///
+  /// Only ever called once the same disposed / generation / exception checks
+  /// that authorise committing those rows have passed. Recording earlier would
+  /// suppress a patch whose fresh row was then discarded or rolled back,
+  /// leaving an older snapshot on screen with nothing to correct it. A
+  /// zero-success pass observed nothing and records nothing.
+  void _recordObservations(
+    ({int watermark, Object epoch})? observation,
+    List<({MediaItem item, String? clientScope})> rows,
+    Set<String> succeededServerIds,
+  ) {
+    final store = _watchStateStore;
+    if (store == null || observation == null || succeededServerIds.isEmpty || rows.isEmpty) return;
+    store.recordObservations(
+      [
+        for (final row in rows)
+          if (succeededServerIds.contains(row.item.serverId)) row,
+      ],
+      watermark: observation.watermark,
+      epoch: observation.epoch,
+    );
+  }
+
   /// Whether the profile binder is still wiring servers — a no-servers load
   /// during binding stays in the loading state instead of flashing an error,
   /// and a zero-success pass during binding stays in the loading state
@@ -100,19 +174,21 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
   String? _errorMessage;
   int _loadGeneration = 0;
   int _contentRevision = 0;
+  int _commitRevision = 0;
+  DiscoverRefreshOutcome _lastOutcome = DiscoverRefreshOutcome.cancelled;
   Future<void>? _continueWatchingRefreshFuture;
   bool _continueWatchingRefreshQueued = false;
 
   Set<String> _lastSeenHiddenKeys = {};
   List<String> _lastSeenLibraryOrderKeys = const [];
 
-  /// Online servers whose Continue Watching fetch succeeded in the current
+  /// Online servers whose Continue Watching legs succeeded without a failure
-  /// on-deck list. Tracked separately from hubs so a transient failure in one
+  /// or cancellation in the current on-deck list. Tracked separately from hubs
-  /// surface does not cache the other as loaded forever or force unnecessary
+  /// so a transient failure in one surface does not cache the other as loaded
-  /// refetches.
+  /// forever or force unnecessary refetches.
   Set<String> _loadedOnDeckServerIds = {};
 
-  /// Online servers whose home-hub fetch succeeded in the current hub list.
+  /// Online servers whose home-hub legs all succeeded in the current hub list.
   Set<String> _loadedHubServerIds = {};
 
   Set<String> get _fullyLoadedServerIds => _loadedOnDeckServerIds.intersection(_loadedHubServerIds);
@@ -167,12 +243,48 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
     return _loadCoordinator.requestFull();
   }
 
+  Future<DiscoverRefreshOutcome> refreshNow() async {
+    if (isDisposed) return DiscoverRefreshOutcome.cancelled;
+    await _loadCoordinator.requestFull();
+    if (isDisposed) return DiscoverRefreshOutcome.cancelled;
+    return _lastOutcome;
+  }
+
   /// Whether a [load] pass is already running. The startup online-entry hook
   /// uses this to skip a prime that would only duplicate the load the screen
   /// started in `initState`.
   bool get isLoadInFlight => _loadCoordinator.isBusy;
 
   Future<void> _loadOnce() async {
+    var outcome = DiscoverRefreshOutcome.cancelled;
+    var passClearedExceptionBoundary = false;
+    List<MediaItem>? systemShelfPassToken;
+    // Observations are staged with the pass, not recorded as each leg lands:
+    // suppressing a patch whose fresh row is then discarded or rolled back
+    // would leave an older snapshot on screen with nothing to correct it.
+    final pendingObservations = <({MediaItem item, String? clientScope})>[];
+    final observedServerIds = <String>{};
+    // Assigned inside the try, after the preparatory awaits, so the watermark
+    // brackets the network calls rather than the whole method.
+    ({int watermark, Object epoch})? observation;
+    final succeededServerIds = <String>{};
+    final failedServerIds = <String>{};
+    final cancelledServerIds = <String>{};
+    final previousOnDeck = _onDeck;
+    final previousHubs = _hubs;
+    final previousHasMoreContinueWatching = _hasMoreContinueWatching;
+    final previousLoadedOnDeckServerIds = _loadedOnDeckServerIds;
+    final previousLoadedHubServerIds = _loadedHubServerIds;
+    final previousOnDeckState = _onDeckState;
+    final previousHubsState = _hubsState;
+    final previousLoadGeneration = _loadGeneration;
+    final previousCommitRevision = _commitRevision;
+    var expectedCommitRevision = previousCommitRevision;
+    var onDeckFetchCompleted = false;
+    var hubFetchCompleted = false;
+    var replacedOnDeck = false;
+
+    try {
       // Yield to the microtask queue before the first notify so a load()
       // kicked off during build (the screen's initState) doesn't mark
       // listening widgets dirty mid-build.
@@ -185,7 +297,6 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       _errorMessage = null;
       safeNotifyListeners();
 
-    try {
       if (!_multiServer.hasConnectedServers) {
         if (isProfileBinding()) return;
         throw Exception('No servers available');
@@ -202,6 +313,11 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
 
       // On-deck and hubs fetch in parallel; on-deck is published as soon as
       // it lands so the hero renders while hubs are still loading.
+      //
+      // The watermark is captured immediately before the requests, after
+      // every preparatory await: a patch recorded later has a higher sequence
+      // and must survive this pass's observations.
+      observation = _beginObservation();
       final onDeckFuture = aggregation.getOnDeckFromAllServers(
         limit: _continueWatchingProbeLimit,
         hiddenLibraryKeys: _hiddenLibraries.hiddenLibraryKeys,
@@ -219,13 +335,23 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       // follow-up load is guaranteed (binding-settle prime, or
       // syncToOnlineServers falling through to load() while not loaded).
       final fetchedOnDeck = await onDeckFuture;
+      onDeckFetchCompleted = true;
+      succeededServerIds.addAll(fetchedOnDeck.succeededServerIds);
+      failedServerIds.addAll(fetchedOnDeck.failedServerIds);
+      cancelledServerIds.addAll(fetchedOnDeck.cancelledServerIds);
+      pendingObservations.addAll(fetchedOnDeck.observedItems);
+      observedServerIds.addAll(fetchedOnDeck.succeededServerIds);
       if (isDisposed) return;
       if (fetchedOnDeck.succeededServerIds.isEmpty && _onDeck.isNotEmpty) {
         // Keep the stale rows; the empty succeeded set makes the next status
         // emission refetch every server.
         appLogger.w('DiscoverProvider: on-deck pass failed on all servers; keeping previous items');
         _onDeckState = DiscoverLoadState.loaded;
-        _loadedOnDeckServerIds = fetchedOnDeck.succeededServerIds;
+        _loadedOnDeckServerIds = _authoritativeSucceededServerIds(
+          fetchedOnDeck.succeededServerIds,
+          fetchedOnDeck.failedServerIds,
+          fetchedOnDeck.cancelledServerIds,
+        );
         safeNotifyListeners();
       } else if (fetchedOnDeck.succeededServerIds.isEmpty &&
           (fetchedOnDeck.cancelledServerIds.isNotEmpty || isProfileBinding())) {
@@ -235,25 +361,60 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
         appLogger.d('DiscoverProvider: on-deck pass disrupted with no prior content; keeping loading state');
       } else {
         _applyOnDeck(fetchedOnDeck.items);
+        ++expectedCommitRevision;
+        replacedOnDeck = true;
         _onDeckState = DiscoverLoadState.loaded;
-        _loadedOnDeckServerIds = fetchedOnDeck.succeededServerIds;
+        _loadedOnDeckServerIds = _authoritativeSucceededServerIds(
-        _loadGeneration++;
+          fetchedOnDeck.succeededServerIds,
+          fetchedOnDeck.failedServerIds,
+          fetchedOnDeck.cancelledServerIds,
+        );
+        systemShelfPassToken = List<MediaItem>.unmodifiable(_onDeck);
         safeNotifyListeners();
-        unawaited(_syncSystemShelf(_onDeck));
       }
 
       final fetchedHubs = await hubsFuture;
+      hubFetchCompleted = true;
+      succeededServerIds.addAll(fetchedHubs.succeededServerIds);
+      failedServerIds.addAll(fetchedHubs.failedServerIds);
+      pendingObservations.addAll(fetchedHubs.observedItems);
+      observedServerIds.addAll(fetchedHubs.succeededServerIds);
+      cancelledServerIds.addAll(fetchedHubs.cancelledServerIds);
       if (isDisposed) return;
 
       if (fetchedHubs.succeededServerIds.isEmpty && _hubs.isNotEmpty) {
         appLogger.w('DiscoverProvider: hub pass failed on all servers; keeping previous hubs');
         _hubsState = DiscoverLoadState.loaded;
-        _loadedHubServerIds = fetchedHubs.succeededServerIds;
+        _loadedHubServerIds = _authoritativeSucceededServerIds(
+          fetchedHubs.succeededServerIds,
+          fetchedHubs.failedServerIds,
+          fetchedHubs.cancelledServerIds,
+        );
         safeNotifyListeners();
+        outcome = _refreshOutcome(
+          succeededServerIds: succeededServerIds,
+          failedServerIds: failedServerIds,
+          cancelledServerIds: cancelledServerIds,
+          cancelled: isProfileBinding(),
+        );
+        if (replacedOnDeck && outcome != DiscoverRefreshOutcome.failed && outcome != DiscoverRefreshOutcome.cancelled) {
+          ++_loadGeneration;
+        }
+        passClearedExceptionBoundary = true;
         return;
       }
       if (fetchedHubs.succeededServerIds.isEmpty && (fetchedHubs.cancelledServerIds.isNotEmpty || isProfileBinding())) {
         appLogger.d('DiscoverProvider: hub pass disrupted with no prior content; keeping loading state');
+        outcome = _refreshOutcome(
+          succeededServerIds: succeededServerIds,
+          failedServerIds: failedServerIds,
+          cancelledServerIds: cancelledServerIds,
+          cancelled: isProfileBinding(),
+        );
+        if (replacedOnDeck && outcome != DiscoverRefreshOutcome.failed && outcome != DiscoverRefreshOutcome.cancelled) {
+          ++_loadGeneration;
+        }
+        passClearedExceptionBoundary = true;
         return;
       }
 
@@ -261,17 +422,80 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       sortMediaHubsByLibraryOrder(filteredHubs, _libraries.libraries);
 
       appLogger.d('DiscoverProvider: ${_onDeck.length} on-deck items, ${filteredHubs.length} hubs');
-      _hubs = filteredHubs;
+      _replaceHubs(filteredHubs);
+      ++expectedCommitRevision;
       _hubsState = DiscoverLoadState.loaded;
-      _loadedHubServerIds = fetchedHubs.succeededServerIds;
+      _loadedHubServerIds = _authoritativeSucceededServerIds(
+        fetchedHubs.succeededServerIds,
+        fetchedHubs.failedServerIds,
+        fetchedHubs.cancelledServerIds,
+      );
+      outcome = _refreshOutcome(
+        succeededServerIds: succeededServerIds,
+        failedServerIds: failedServerIds,
+        cancelledServerIds: cancelledServerIds,
+        cancelled: isProfileBinding(),
+      );
+      if (replacedOnDeck && outcome != DiscoverRefreshOutcome.failed && outcome != DiscoverRefreshOutcome.cancelled) {
+        ++_loadGeneration;
+      }
+      passClearedExceptionBoundary = true;
       safeNotifyListeners();
     } catch (e) {
       if (isDisposed) return;
+      outcome = isProfileBinding() ? DiscoverRefreshOutcome.cancelled : DiscoverRefreshOutcome.failed;
       appLogger.e('Failed to load discover content', error: e);
+      final hadPriorContent = previousOnDeck.isNotEmpty || previousHubs.isNotEmpty;
+      if (!hadPriorContent) {
         _errorMessage = e.toString();
         _onDeckState = DiscoverLoadState.error;
         _hubsState = DiscoverLoadState.error;
         safeNotifyListeners();
+        return;
+      }
+
+      if (_commitRevision == expectedCommitRevision) {
+        _replaceOnDeck(
+          _withoutHiddenLibraries(previousOnDeck, _hiddenLibraries.hiddenLibraryKeys),
+          hasMore: previousHasMoreContinueWatching,
+        );
+        _replaceHubs(_hubsWithoutHiddenLibraries(previousHubs, _hiddenLibraries.hiddenLibraryKeys));
+        _loadedOnDeckServerIds = Set<String>.of(previousLoadedOnDeckServerIds);
+        _loadedHubServerIds = Set<String>.of(previousLoadedHubServerIds);
+        _onDeckState = previousOnDeckState;
+        _hubsState = previousHubsState;
+        _loadGeneration = previousLoadGeneration;
+      } else {
+        _filterCurrentContentForHiddenLibraries();
+      }
+
+      final hiddenServerIds = _serverIdsForLibraryKeys(_hiddenLibraries.hiddenLibraryKeys);
+      if (!onDeckFetchCompleted) {
+        _loadedOnDeckServerIds = {};
+      } else {
+        _loadedOnDeckServerIds = Set<String>.of(_loadedOnDeckServerIds)
+          ..removeAll(failedServerIds)
+          ..removeAll(cancelledServerIds);
+      }
+      if (!hubFetchCompleted) {
+        _loadedHubServerIds = {};
+      } else {
+        _loadedHubServerIds = Set<String>.of(_loadedHubServerIds)
+          ..removeAll(failedServerIds)
+          ..removeAll(cancelledServerIds);
+      }
+      _loadedOnDeckServerIds = Set<String>.of(_loadedOnDeckServerIds)..removeAll(hiddenServerIds);
+      _loadedHubServerIds = Set<String>.of(_loadedHubServerIds)..removeAll(hiddenServerIds);
+      _errorMessage = null;
+      _onDeckState = DiscoverLoadState.loaded;
+      _hubsState = DiscoverLoadState.loaded;
+      safeNotifyListeners();
+    } finally {
+      _lastOutcome = outcome;
+      if (passClearedExceptionBoundary && _commitRevision == expectedCommitRevision && !isDisposed) {
+        _recordObservations(observation, pendingObservations, observedServerIds);
+        if (systemShelfPassToken != null) unawaited(_syncSystemShelf(systemShelfPassToken));
+      }
     }
   }
 
@@ -297,6 +521,7 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       final useGlobalHubs = settings.read(SettingsService.useGlobalHubs);
       final aggregation = _multiServer.aggregationService;
 
+      final observation = _beginObservation();
       final Future<OnDeckAggregationResult?> onDeckFuture = onDeckIds.isEmpty
           ? Future<OnDeckAggregationResult?>.value()
           : aggregation.getOnDeckFromAllServers(
@@ -316,6 +541,11 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       final freshOnDeck = await onDeckFuture;
       final freshHubs = await hubsFuture;
       if (isDisposed) return;
+      // Staged, not recorded here: the merge below can still throw, and the
+      // catch keeps the previous rows. Suppressing patches against rows that
+      // were then discarded would strand an older snapshot on screen.
+      final pendingObservations = <({MediaItem item, String? clientScope})>[];
+      final observedServerIds = <String>{};
 
       if (freshOnDeck != null) {
         final hadMore = _hasMoreContinueWatching;
@@ -329,7 +559,9 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
         // The stored list is already trimmed, so the merge can't see old items
         // past the cap — a previously-true "more" affordance stays true.
         if (hadMore) _hasMoreContinueWatching = true;
-        _loadedOnDeckServerIds = {..._loadedOnDeckServerIds, ...freshOnDeck.succeededServerIds};
+        _loadedOnDeckServerIds = {..._loadedOnDeckServerIds, ...freshOnDeck.succeededServerIds}
+          ..removeAll(freshOnDeck.failedServerIds)
+          ..removeAll(freshOnDeck.cancelledServerIds);
         // No _loadGeneration bump: a delta behaves like the background Continue
         // Watching refresh (the hero clamps instead of resetting).
       }
@@ -341,10 +573,22 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
           ..._filterDiscoverHubs(freshHubs.hubs),
         ];
         sortMediaHubsByLibraryOrder(mergedHubs, _libraries.libraries);
-        _hubs = mergedHubs;
+        _replaceHubs(mergedHubs);
-        _loadedHubServerIds = {..._loadedHubServerIds, ...succeededHubIds};
+        _loadedHubServerIds = {..._loadedHubServerIds, ...succeededHubIds}
+          ..removeAll(freshHubs.failedServerIds)
+          ..removeAll(freshHubs.cancelledServerIds);
       }
 
+      if (freshOnDeck != null) {
+        pendingObservations.addAll(freshOnDeck.observedItems);
+        observedServerIds.addAll(freshOnDeck.succeededServerIds);
+      }
+      if (freshHubs != null) {
+        pendingObservations.addAll(freshHubs.observedItems);
+        observedServerIds.addAll(freshHubs.succeededServerIds);
+      }
+      _recordObservations(observation, pendingObservations, observedServerIds);
+
       appLogger.d('DiscoverProvider: ${_onDeck.length} on-deck items, ${_hubs.length} hubs after merging $ids');
       safeNotifyListeners();
       unawaited(_syncSystemShelf(_onDeck));
@@ -369,6 +613,69 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
     }).toList();
   }
 
+  Set<String> _authoritativeSucceededServerIds(
+    Set<String> succeededServerIds,
+    Set<String> failedServerIds,
+    Set<String> cancelledServerIds,
+  ) {
+    return Set<String>.of(succeededServerIds)
+      ..removeAll(failedServerIds)
+      ..removeAll(cancelledServerIds);
+  }
+
+  List<MediaItem> _withoutHiddenLibraries(List<MediaItem> items, Set<String> hiddenLibraryKeys) {
+    if (hiddenLibraryKeys.isEmpty) return items;
+    return items.where((item) {
+      final libraryKey = item.libraryGlobalKey;
+      return libraryKey == null || !hiddenLibraryKeys.contains(libraryKey);
+    }).toList();
+  }
+
+  List<MediaHub> _hubsWithoutHiddenLibraries(List<MediaHub> hubs, Set<String> hiddenLibraryKeys) {
+    if (hiddenLibraryKeys.isEmpty) return hubs;
+    final filteredHubs = <MediaHub>[];
+    for (final hub in hubs) {
+      final filteredItems = hub.items.where((item) {
+        var libraryKey = item.libraryGlobalKey;
+        final libraryId = item.libraryId;
+        final serverId = item.serverId ?? hub.serverId;
+        if (libraryKey == null && libraryId != null && serverId != null) {
+          libraryKey = buildGlobalKey(ServerId(serverId), libraryId);
+        }
+        return libraryKey == null || !hiddenLibraryKeys.contains(libraryKey);
+      }).toList();
+      if (filteredItems.isNotEmpty) {
+        filteredHubs.add(
+          filteredItems.length == hub.items.length
+              ? hub
+              : hub.copyWith(items: filteredItems, size: filteredItems.length),
+        );
+      }
+    }
+    return filteredHubs;
+  }
+
+  Set<String> _serverIdsForLibraryKeys(Set<String> libraryKeys) {
+    final serverIds = <String>{};
+    for (final key in libraryKeys) {
+      final parsed = parseGlobalKey(key);
+      if (parsed != null) serverIds.add(parsed.serverId);
+    }
+    return serverIds;
+  }
+
+  void _filterCurrentContentForHiddenLibraries() {
+    final hiddenLibraryKeys = _hiddenLibraries.hiddenLibraryKeys;
+    final filteredOnDeck = _withoutHiddenLibraries(_onDeck, hiddenLibraryKeys);
+    if (filteredOnDeck.length != _onDeck.length) {
+      _replaceOnDeck(filteredOnDeck, hasMore: _hasMoreContinueWatching);
+    }
+    final filteredHubs = _hubsWithoutHiddenLibraries(_hubs, hiddenLibraryKeys);
+    if (!listEquals(filteredHubs, _hubs)) {
+      _replaceHubs(filteredHubs);
+    }
+  }
+
   /// Background refresh of Continue Watching only. Concurrent events coalesce
   /// into the active request plus at most one trailing fresh request.
   Future<void> refreshContinueWatching() {
@@ -399,17 +706,30 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       if (!_multiServer.hasConnectedServers) return;
       final revision = _contentRevision;
       final hiddenKeys = Set<String>.of(_hiddenLibraries.hiddenLibraryKeys);
+      final observation = _beginObservation();
       final fetched = await _multiServer.aggregationService.getOnDeckFromAllServers(
         limit: _continueWatchingProbeLimit,
         hiddenLibraryKeys: hiddenKeys,
       );
       if (isDisposed) return;
       if (revision != _contentRevision) {
+        // A newer mutation landed while this was in flight, so these rows are
+        // discarded — recording them would suppress patches against data the
+        // user never sees.
         _continueWatchingRefreshQueued = true;
         return;
       }
+      _loadedOnDeckServerIds = _authoritativeSucceededServerIds(
+        fetched.succeededServerIds,
+        fetched.failedServerIds,
+        fetched.cancelledServerIds,
+      );
+      if (fetched.succeededServerIds.isEmpty) {
+        safeNotifyListeners();
+        return;
+      }
       _applyOnDeck(fetched.items);
-      _loadedOnDeckServerIds = fetched.succeededServerIds;
+      _recordObservations(observation, fetched.observedItems, fetched.succeededServerIds);
       safeNotifyListeners();
       unawaited(_syncSystemShelf(_onDeck));
     } catch (e) {
@@ -422,9 +742,13 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
     if (!_multiServer.hasConnectedServers) return const [];
     await _hiddenLibraries.ensureInitialized();
     if (isDisposed) return const [];
+    final observation = _beginObservation();
     final fetched = await _multiServer.aggregationService.getOnDeckFromAllServers(
       hiddenLibraryKeys: _hiddenLibraries.hiddenLibraryKeys,
     );
+    // "View All" renders through the same overlay as the row it expands, so
+    // it has to reconcile too or the stale patch simply reappears there.
+    _recordObservations(observation, fetched.observedItems, fetched.succeededServerIds);
     return fetched.items;
   }
 
@@ -448,7 +772,7 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
   void _updateItemInLists(String sourceGlobalKey, MediaItem updatedItem) {
     final onDeckIndex = _onDeck.indexWhere((item) => item.globalKey == sourceGlobalKey);
     if (onDeckIndex != -1) {
-      _onDeck = List.of(_onDeck)..[onDeckIndex] = updatedItem;
+      _replaceOnDeck(List.of(_onDeck)..[onDeckIndex] = updatedItem, hasMore: _hasMoreContinueWatching);
     }
 
     for (var i = 0; i < _hubs.length; i++) {
@@ -457,15 +781,25 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       if (itemIndex != -1) {
         final newItems = List<MediaItem>.from(hub.items);
         newItems[itemIndex] = updatedItem;
-        _hubs = List.of(_hubs)..[i] = hub.copyWith(items: newItems);
+        _replaceHubs(List.of(_hubs)..[i] = hub.copyWith(items: newItems));
       }
     }
   }
 
   void _applyOnDeck(List<MediaItem> fetched) {
     final hasMore = fetched.length > continueWatchingPreviewLimit;
-    _onDeck = hasMore ? fetched.take(continueWatchingPreviewLimit).toList() : fetched;
+    _replaceOnDeck(hasMore ? fetched.take(continueWatchingPreviewLimit).toList() : fetched, hasMore: hasMore);
+  }
+
+  void _replaceOnDeck(List<MediaItem> onDeck, {required bool hasMore}) {
+    _onDeck = onDeck;
     _hasMoreContinueWatching = hasMore;
+    ++_commitRevision;
+  }
+
+  void _replaceHubs(List<MediaHub> hubs) {
+    _hubs = hubs;
+    ++_commitRevision;
   }
 
   // --- Event reactions -----------------------------------------------------
@@ -481,7 +815,10 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
       final viewOffset = event.viewOffset;
       final index = _onDeck.indexWhere((item) => item.globalKey == event.globalKey);
       if (viewOffset != null && index != -1 && _onDeck[index].viewOffsetMs != viewOffset) {
-        _onDeck = List.of(_onDeck)..[index] = _onDeck[index].copyWith(viewOffsetMs: viewOffset);
+        _replaceOnDeck(
+          List.of(_onDeck)..[index] = _onDeck[index].copyWith(viewOffsetMs: viewOffset),
+          hasMore: _hasMoreContinueWatching,
+        );
         safeNotifyListeners();
         unawaited(_syncSystemShelf(_onDeck));
       }
@@ -507,7 +844,7 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
   void _evictFromOnDeck(bool Function(MediaItem item) matches) {
     final remaining = _onDeck.where((item) => !matches(item)).toList();
     if (remaining.length == _onDeck.length) return;
-    _onDeck = remaining;
+    _replaceOnDeck(remaining, hasMore: _hasMoreContinueWatching);
     safeNotifyListeners();
     unawaited(_syncSystemShelf(_onDeck));
   }
@@ -533,14 +870,14 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
     var changed = false;
     final remainingOnDeck = _onDeck.where((item) => !affected(item)).toList();
     if (remainingOnDeck.length != _onDeck.length) {
-      _onDeck = remainingOnDeck;
+      _replaceOnDeck(remainingOnDeck, hasMore: _hasMoreContinueWatching);
       changed = true;
     }
     for (var i = 0; i < _hubs.length; i++) {
       final hub = _hubs[i];
       final newItems = hub.items.where((item) => !affected(item)).toList();
       if (newItems.length != hub.items.length) {
-        _hubs = List.of(_hubs)..[i] = hub.copyWith(items: newItems);
+        _replaceHubs(List.of(_hubs)..[i] = hub.copyWith(items: newItems));
         changed = true;
       }
     }
@@ -565,7 +902,7 @@ class DiscoverProvider extends ChangeNotifier with DisposableChangeNotifierMixin
 
     final sortedHubs = List<MediaHub>.from(_hubs);
     if (!sortMediaHubsByLibraryOrder(sortedHubs, _libraries.libraries)) return;
-    _hubs = sortedHubs;
+    _replaceHubs(sortedHubs);
     safeNotifyListeners();
   }
 

lib/providers/offline_watch_provider.dart

@@ -144,12 +144,18 @@ class OfflineWatchProvider extends ChangeNotifier with DisposableChangeNotifierM
     required String itemId,
     required bool isNowWatched,
     required WatchStateChangeType changeType,
+    required WatchPatchId patchId,
     String? cacheServerId,
   }) {
     final globalKey = buildGlobalKey(ServerId(serverId), itemId);
     final metadata = _downloadProvider.getMetadata(globalKey);
     if (metadata != null) {
-      WatchStateNotifier().notifyWatched(item: metadata, isNowWatched: isNowWatched, cacheServerId: cacheServerId);
+      WatchStateNotifier().notifyWatched(
+        item: metadata,
+        isNowWatched: isNowWatched,
+        cacheServerId: cacheServerId,
+        patchId: patchId,
+      );
     } else {
       // Fallback: emit minimal event without parent chain.
       WatchStateNotifier().notify(
@@ -161,6 +167,7 @@ class OfflineWatchProvider extends ChangeNotifier with DisposableChangeNotifierM
           parentChain: [],
           mediaType: 'unknown',
           isNowWatched: isNowWatched,
+          patchId: patchId,
         ),
       );
     }
@@ -170,13 +177,14 @@ class OfflineWatchProvider extends ChangeNotifier with DisposableChangeNotifierM
   ///
   /// This queues the action for sync when online and emits a [WatchStateEvent].
   Future<void> markAsWatched({required ServerId serverId, required String itemId}) async {
-    final cacheServerId = await _syncService.queueMarkWatched(serverId: serverId, itemId: itemId);
+    final queued = await _syncService.queueMarkWatched(serverId: serverId, itemId: itemId);
     _emitWatchStateChange(
       serverId: serverId,
       itemId: itemId,
       isNowWatched: true,
       changeType: WatchStateChangeType.watched,
-      cacheServerId: cacheServerId,
+      cacheServerId: queued.clientScopeId,
+      patchId: WatchPatchId.offlineAction(profileId: queued.profileId, rowId: queued.rowId, revision: queued.revision),
     );
     safeNotifyListeners();
     _autoDeleteIfWatched(serverId, itemId);
@@ -212,13 +220,14 @@ class OfflineWatchProvider extends ChangeNotifier with DisposableChangeNotifierM
   ///
   /// This queues the action for sync when online and emits a [WatchStateEvent].
   Future<void> markAsUnwatched({required ServerId serverId, required String itemId}) async {
-    final cacheServerId = await _syncService.queueMarkUnwatched(serverId: serverId, itemId: itemId);
+    final queued = await _syncService.queueMarkUnwatched(serverId: serverId, itemId: itemId);
     _emitWatchStateChange(
       serverId: serverId,
       itemId: itemId,
       isNowWatched: false,
       changeType: WatchStateChangeType.unwatched,
-      cacheServerId: cacheServerId,
+      cacheServerId: queued.clientScopeId,
+      patchId: WatchPatchId.offlineAction(profileId: queued.profileId, rowId: queued.rowId, revision: queued.revision),
     );
     safeNotifyListeners();
   }

lib/providers/watch_state_store.dart

@@ -26,19 +26,57 @@ class HydratedWatchStatePatch {
   });
 }
 
+/// Key for [WatchStateStore]'s observation map: a global key already resolved
+/// through the client scope its request was issued under. A plain `String`
+/// would work, but the wrapper keeps the two key spaces from being confused
+/// with the patch maps' keys at a glance.
+class _ObservationKey {
+  final String value;
+
+  const _ObservationKey(this.value);
+
+  @override
+  bool operator ==(Object other) => other is _ObservationKey && other.value == value;
+
+  @override
+  int get hashCode => value.hashCode;
+}
+
 class _WatchStatePatchEntry {
   final WatchStateSnapshot patch;
   final int updatedAt;
   final int sequence;
   final bool isSessionEvent;
 
+  /// Whether the server had accepted this exact state when the entry was
+  /// recorded. Only an acknowledged entry may be superseded by a later
+  /// authoritative read; an unacknowledged one is a write still owed to the
+  /// server and must outlive any read (#1829).
+  final bool serverAcknowledged;
+
+  /// Identity used to promote this exact entry once its write settles.
+  /// Promotion never matches by global key: a same-item rewatch may already
+  /// have replaced the entry, and it must not be promoted in its place.
+  final WatchPatchId? patchId;
+
   const _WatchStatePatchEntry(
     this.patch, {
     required this.updatedAt,
     required this.sequence,
     required this.isSessionEvent,
+    this.serverAcknowledged = false,
+    this.patchId,
   });
 
+  _WatchStatePatchEntry acknowledgedAt(int sequence) => _WatchStatePatchEntry(
+    patch,
+    updatedAt: updatedAt,
+    sequence: sequence,
+    isSessionEvent: true,
+    serverAcknowledged: true,
+    patchId: patchId,
+  );
+
   bool isNewerThan(_WatchStatePatchEntry other) {
     if (updatedAt != other.updatedAt) return updatedAt > other.updatedAt;
     if (isSessionEvent != other.isSessionEvent) return isSessionEvent;
@@ -51,10 +89,12 @@ class _WatchStatePatchEntry {
       other.patch == patch &&
       other.updatedAt == updatedAt &&
       other.sequence == sequence &&
-      other.isSessionEvent == isSessionEvent;
+      other.isSessionEvent == isSessionEvent &&
+      other.serverAcknowledged == serverAcknowledged &&
+      other.patchId == patchId;
 
   @override
-  int get hashCode => Object.hash(patch, updatedAt, sequence, isSessionEvent);
+  int get hashCode => Object.hash(patch, updatedAt, sequence, isSessionEvent, serverAcknowledged, patchId);
 }
 
 /// The single session-local layer for watch-state freshness.
@@ -69,50 +109,175 @@ class _WatchStatePatchEntry {
 class WatchStateStore extends ChangeNotifier with DisposableChangeNotifierMixin {
   WatchStateStore() {
     _subscription = WatchStateNotifier().stream.listen(_onWatchStateEvent);
+    // A second, separately owned subscription: promotions are deliberately
+    // not WatchStateEvents, and BaseNotifier carries exactly one typed
+    // stream, so they cannot arrive on the channel above.
+    _promotionSubscription = WatchPatchPromotionNotifier().stream.listen(_onPromotion);
   }
 
   StreamSubscription<WatchStateEvent>? _subscription;
+  StreamSubscription<WatchPatchPromotion>? _promotionSubscription;
   final Map<String, _WatchStatePatchEntry> _patches = {};
   final Map<String, _WatchStatePatchEntry> _hydratedPatches = {};
   String? _activeProfileId;
   Map<String, String?> _activeClientScopesByServer = const {};
   int _sequence = 0;
 
-  _WatchStatePatchEntry? _exactEntryFor(String globalKey) {
+  /// Highest watermark at which a qualifying authoritative response returned a
-    final session = _patches[globalKey];
+  /// key, under the client scope its request was issued for.
+  ///
+  /// This records *that* the server re-observed an item, never *what* it
+  /// observed. Storing the observed state instead would need full fidelity
+  /// ([WatchStateSnapshot] cannot hold a container's leaf counts) and would
+  /// make two responses sharing a watermark first-response-wins. A watermark
+  /// has neither problem: it is a single int, and [recordObservations] keeps
+  /// the max, so response order cannot matter.
+  final Map<_ObservationKey, int> _observedAt = {};
+
+  /// Snapshot the current sequence *before* issuing an authoritative request;
+  /// pass it back to [recordObservations] when that request succeeds.
+  int get observationWatermark => _sequence;
+
+  /// Identifies this store instance. Watermarks are store-local, but clients —
+  /// and their in-flight requests — outlive a profile switch, so a response
+  /// minted against the previous store must not be applied to this one.
+  final Object _epoch = Object();
+
+  Object get observationEpoch => _epoch;
+
+  _WatchStatePatchEntry? _exactEntryFor(String globalKey, int? observedAt) {
+    final session = _live(_patches[globalKey], observedAt);
     final hydrated = _hydratedPatches[globalKey];
     if (session == null) return hydrated;
     if (hydrated == null) return session;
     return session.isNewerThan(hydrated) ? session : hydrated;
   }
 
-  _WatchStatePatchEntry? _entryFor(String globalKey) {
+  /// Drops an acknowledged session entry the server has re-observed since.
+  ///
+  /// Only acknowledged entries are eligible: an unacknowledged one is a write
+  /// still owed to the server, and a read must never retire it. Hydrated
+  /// entries are the persisted owed-write layer and are likewise untouched —
+  /// filtering happens per layer, before [_exactEntryFor] chooses, so a
+  /// suppressed session entry falls back to an older hydrated one instead of
+  /// suppressing both.
+  static _WatchStatePatchEntry? _live(_WatchStatePatchEntry? entry, int? observedAt) {
+    if (entry == null) return null;
+    if (!entry.serverAcknowledged || observedAt == null) return entry;
+    return observedAt >= entry.sequence ? null : entry;
+  }
+
+  /// Candidate keys for [globalKey], in the order [_entryFor] consults them:
+  /// the active client scope first, then the public fallback.
+  List<String> _candidateKeys(String globalKey) {
     final parsed = parseGlobalKey(globalKey);
     if (parsed != null) {
       final scoped = _activeClientScopesByServer[parsed.serverId];
       if (scoped != null && scoped.isNotEmpty) {
-        return _exactEntryFor(buildGlobalKey(ServerId(scoped), parsed.ratingKey)) ?? _exactEntryFor(globalKey);
+        return [buildGlobalKey(ServerId(scoped), parsed.ratingKey), globalKey];
       }
     }
-    return _exactEntryFor(globalKey);
+    return [globalKey];
   }
 
+  _WatchStatePatchEntry? _entryFor(String globalKey, [int? observedAt]) {
+    for (final key in _candidateKeys(globalKey)) {
+      final entry = _exactEntryFor(key, observedAt);
+      if (entry != null) return entry;
+    }
+    return null;
+  }
+
+  /// The watermark at which [globalKey] was last authoritatively observed.
+  ///
+  /// Resolved through the *primary* candidate only — the active client scope
+  /// when there is one, else the public key. Deliberately not a max over both:
+  /// on a user-scoped backend the public server id is shared between users, so
+  /// letting a public observation satisfy a scoped patch would suppress
+  /// another user's watch state.
+  int? _observationFor(String globalKey) => _observedAt[_ObservationKey(_candidateKeys(globalKey).first)];
+
   @visibleForTesting
-  WatchStateSnapshot? patchForGlobalKey(String globalKey) => _entryFor(globalKey)?.patch;
+  WatchStateSnapshot? patchForGlobalKey(String globalKey) => _entryFor(globalKey, _observationFor(globalKey))?.patch;
 
   WatchStateSnapshot? patchForItem(MediaItem item) {
-    var best = _entryFor(item.globalKey);
+    if (_patches.isEmpty && _hydratedPatches.isEmpty) return null;
+    // One barrier for the whole resolution: an authoritative read of *this*
+    // item already incorporates any container mark that preceded it, so the
+    // ancestor candidates are judged against the item's own observation too.
+    // Suppressing only the item's own entry would let an older season patch
+    // win and render something worse than either value (#1829 review).
+    final observedAt = _observationFor(item.globalKey);
+    var best = _entryFor(item.globalKey, observedAt);
     if (item.parentChain.isNotEmpty) {
       final serverId = serverIdOrNull(item.serverId);
       for (final parentId in item.parentChain) {
         // Mirror MediaItem.globalKey's bare-id fallback when serverId is missing.
-        final entry = _entryFor(serverId != null ? buildGlobalKey(serverId, parentId) : parentId);
+        final entry = _entryFor(serverId != null ? buildGlobalKey(serverId, parentId) : parentId, observedAt);
         if (entry != null && (best == null || entry.isNewerThan(best))) best = entry;
       }
     }
     return best?.patch;
   }
 
+  /// Record that a successful authoritative response returned [rows].
+  ///
+  /// [watermark] must be [observationWatermark] captured immediately before
+  /// the network call, and [epoch] the [observationEpoch] of the store that
+  /// captured it. Entries recorded *during* the request keep a higher
+  /// sequence and survive, so a local action taken mid-flight always wins.
+  ///
+  /// Each row carries the immutable cache scope of the client that fetched
+  /// it: on a user-scoped backend the public server id is shared, so a row
+  /// must only ever reconcile the scope it actually came from.
+  void recordObservations(
+    Iterable<({MediaItem item, String? clientScope})> rows, {
+    required int watermark,
+    required Object epoch,
+  }) {
+    // Watermarks are store-local, but clients — and their in-flight requests
+    // — outlive a profile switch, so a response minted against the previous
+    // store carries a sequence that means nothing here.
+    if (!identical(epoch, _epoch)) return;
+    if (_patches.isEmpty) return;
+    var changed = false;
+    for (final row in rows) {
+      final item = row.item;
+      // An observation is only ever consulted to suppress a patch, so keep
+      // one only while some patch could still apply to this item — its own,
+      // or an ancestor's, which the barrier in [patchForItem] also judges
+      // against this key. That bounds the map by the items the user acted
+      // on rather than by everything ever fetched.
+      if (!_hasSuppressibleEntry(item)) continue;
+      final scope = row.clientScope;
+      final key = _ObservationKey(
+        scope != null && scope.isNotEmpty && scope != item.serverId
+            ? buildGlobalKey(ServerId(scope), item.id)
+            : item.globalKey,
+      );
+      final existing = _observedAt[key];
+      if (existing != null && existing >= watermark) continue;
+      _observedAt[key] = watermark;
+      changed = true;
+    }
+    if (changed) safeNotifyListeners();
+  }
+
+  bool _hasSuppressibleEntry(MediaItem item) {
+    for (final key in _candidateKeys(item.globalKey)) {
+      if (_patches[key]?.serverAcknowledged ?? false) return true;
+    }
+    if (item.parentChain.isEmpty) return false;
+    final serverId = serverIdOrNull(item.serverId);
+    for (final parentId in item.parentChain) {
+      final parentKey = serverId != null ? buildGlobalKey(serverId, parentId) : parentId;
+      for (final key in _candidateKeys(parentKey)) {
+        if (_patches[key]?.serverAcknowledged ?? false) return true;
+      }
+    }
+    return false;
+  }
+
   MediaItem apply(MediaItem item) {
     return applyPatch(item, patchForItem(item));
   }
@@ -127,9 +292,11 @@ class WatchStateStore extends ChangeNotifier with DisposableChangeNotifierMixin
   void setActiveProfileId(String? profileId) {
     if (_activeProfileId == profileId) return;
     _activeProfileId = profileId;
-    if (_patches.isEmpty && _hydratedPatches.isEmpty) return;
+    if (_patches.isEmpty && _hydratedPatches.isEmpty && _observedAt.isEmpty) return;
     _patches.clear();
     _hydratedPatches.clear();
+    // Observations are only meaningful against the patches they suppress.
+    _observedAt.clear();
     safeNotifyListeners();
   }
 
@@ -188,14 +355,56 @@ class WatchStateStore extends ChangeNotifier with DisposableChangeNotifierMixin
       updatedAt: DateTime.now().millisecondsSinceEpoch,
       sequence: ++_sequence,
       isSessionEvent: true,
+      serverAcknowledged: event.serverAcknowledged,
+      patchId: event.patchId,
     );
     safeNotifyListeners();
   }
 
+  /// The write behind [promotion] has settled, so its entry may now be
+  /// superseded by a later authoritative read.
+  ///
+  /// Matching is by patch identity, never by key: an unknown id — because a
+  /// newer action replaced the entry, or because the process restarted — is a
+  /// deliberate no-op. The fresh sequence matters as much as the flag: an
+  /// observation captured *before* the write landed must not suppress the
+  /// entry it predates.
+  void _onPromotion(WatchPatchPromotion promotion) {
+    final sessionKey = _keyForPatchId(_patches, promotion.patchId);
+    if (sessionKey != null) {
+      _patches[sessionKey] = _patches[sessionKey]!.acknowledgedAt(++_sequence);
+      safeNotifyListeners();
+      return;
+    }
+
+    // A hydrated entry is the persisted owed-write layer. Once the write has
+    // landed it no longer belongs there, and leaving it would make it
+    // immortal, since suppression only ever considers session entries. Move
+    // it — but never over a newer session entry on the same key, which would
+    // defeat the exact-identity rule promotion exists for.
+    final hydratedKey = _keyForPatchId(_hydratedPatches, promotion.patchId);
+    if (hydratedKey == null) return;
+    final promoted = _hydratedPatches.remove(hydratedKey)!;
+    final existing = _patches[hydratedKey];
+    if (existing == null || existing.patchId == promotion.patchId) {
+      _patches[hydratedKey] = promoted.acknowledgedAt(++_sequence);
+    }
+    safeNotifyListeners();
+  }
+
+  static String? _keyForPatchId(Map<String, _WatchStatePatchEntry> entries, WatchPatchId patchId) {
+    for (final entry in entries.entries) {
+      if (entry.value.patchId == patchId) return entry.key;
+    }
+    return null;
+  }
+
   @override
   void dispose() {
     _subscription?.cancel();
     _subscription = null;
+    _promotionSubscription?.cancel();
+    _promotionSubscription = null;
     super.dispose();
   }
 }

lib/screens/discover_screen.dart

@@ -50,6 +50,7 @@ import '../utils/formatters.dart';
 import '../utils/hub_icons.dart';
 import '../utils/media_navigation_helper.dart';
 import '../utils/provider_extensions.dart';
+import '../utils/snackbar_helper.dart';
 import '../utils/video_player_navigation.dart';
 import '../utils/layout_constants.dart';
 import '../utils/platform_detector.dart';
@@ -756,7 +757,23 @@ class _DiscoverScreenState extends State<DiscoverScreen>
                 onNavigateLeft: _navigateToSidebar,
                 onNavigateDown: _focusContentFromAppBar,
                 actions: [
-                  FocusableAction(icon: Symbols.refresh_rounded, iconColor: foregroundColor, onPressed: _discover.load),
+                  FocusableAction(
+                    icon: Symbols.refresh_rounded,
+                    iconColor: foregroundColor,
+                    onPressed: () async {
+                      final outcome = await _discover.refreshNow();
+                      if (!context.mounted) return;
+                      switch (outcome) {
+                        case DiscoverRefreshOutcome.failed:
+                          showErrorSnackBar(context, t.errors.unableToLoad(context: t.discover.title));
+                        case DiscoverRefreshOutcome.degraded:
+                          appLogger.w('Discover refresh completed with partial server failures');
+                        case DiscoverRefreshOutcome.cancelled:
+                        case DiscoverRefreshOutcome.refreshed:
+                          break;
+                      }
+                    },
+                  ),
                   // Watch Together
                   FocusableAction(
                     onPressed: () =>

lib/services/data_aggregation_service.dart

@@ -16,16 +16,38 @@ import '../utils/media_server_http_client.dart';
 import 'local_playback_history.dart';
 import 'multi_server_manager.dart';
 
+/// A row a successful response actually returned, paired with the immutable
+/// cache scope of the client that fetched it.
+///
+/// The scope cannot be recovered later: flattening and dedup lose which
+/// client produced a row, and a user-scoped backend's public server id is
+/// shared between users — so reconciling by public key could suppress a
+/// different user's watch state.
+typedef ObservedRow = ({MediaItem item, String? clientScope});
+
+/// [observedItems] are the rows the successful responses actually returned,
+/// *before* dedup or limiting. Continue Watching dedup can drop one server's
+/// copy in favour of another's, and a dropped copy would otherwise leave that
+/// key unreconciled even though its server did return it (#1829).
 typedef OnDeckAggregationResult = ({
   List<MediaItem> items,
+  List<ObservedRow> observedItems,
   Set<String> succeededServerIds,
   Set<String> cancelledServerIds,
+  Set<String> failedServerIds,
+});
+typedef HubAggregationResult = ({
+  List<MediaHub> hubs,
+  List<ObservedRow> observedItems,
+  Set<String> succeededServerIds,
+  Set<String> cancelledServerIds,
+  Set<String> failedServerIds,
 });
-typedef HubAggregationResult = ({List<MediaHub> hubs, Set<String> succeededServerIds, Set<String> cancelledServerIds});
 typedef LibraryAggregationResult = ({
   List<MediaLibrary> libraries,
   Set<String> succeededServerIds,
   Set<String> cancelledServerIds,
+  Set<String> failedServerIds,
 });
 typedef SearchAggregationResult = ({
   List<MediaItem> items,
@@ -165,6 +187,7 @@ class DataAggregationService {
         libraries: const <MediaLibrary>[],
         succeededServerIds: const <String>{},
         cancelledServerIds: const <String>{},
+        failedServerIds: const <String>{},
       );
     }
     final fetched = await _fanOut<MediaLibrary>(
@@ -176,6 +199,7 @@ class DataAggregationService {
       libraries: fetched.items,
       succeededServerIds: fetched.succeededServerIds,
       cancelledServerIds: fetched.cancelledServerIds,
+      failedServerIds: fetched.failedServerIds,
     );
   }
 
@@ -190,14 +214,27 @@ class DataAggregationService {
   }) async {
     final clients = _clientsFor(serverIds);
     if (clients.isEmpty) {
-      appLogger.w('No online servers available for fetching on deck');
+      return (
-      return (items: const <MediaItem>[], succeededServerIds: const <String>{}, cancelledServerIds: const <String>{});
+        items: const <MediaItem>[],
+        observedItems: const <ObservedRow>[],
+        succeededServerIds: const <String>{},
+        cancelledServerIds: const <String>{},
+        failedServerIds: const <String>{},
+      );
     }
 
+    final observedRows = <ObservedRow>[];
     final fetched = await _fanOut<MediaItem>(
       clients,
       failureMessage: (serverId) => 'Failed on-deck fetch from $serverId',
-      fetch: (_, client) => client.fetchContinueWatching(count: limit),
+      fetch: (_, client) async {
+        final rows = await client.fetchContinueWatching(count: limit);
+        // Capture the scope here: after the fan-out flattens and dedup runs,
+        // there is no way back to the client that produced a row.
+        final scope = client.cacheServerId;
+        observedRows.addAll([for (final row in rows) (item: row, clientScope: scope)]);
+        return rows;
+      },
     );
     // Filter out items from hidden libraries
     var filteredOnDeck = _withoutHiddenLibraries(fetched.items, hiddenLibraryKeys);
@@ -216,8 +253,12 @@ class DataAggregationService {
 
     return (
       items: items,
+      // Pre-dedup, pre-limit: a copy dropped by dedup or trimmed by the limit
+      // was still authoritatively returned by its server.
+      observedItems: observedRows,
       succeededServerIds: fetched.succeededServerIds,
       cancelledServerIds: fetched.cancelledServerIds,
+      failedServerIds: fetched.failedServerIds,
     );
   }
 
@@ -436,7 +477,13 @@ class DataAggregationService {
     final clients = _clientsFor(serverIds);
     if (clients.isEmpty) {
       appLogger.w('No online servers available for fetching hubs');
-      return (hubs: const <MediaHub>[], succeededServerIds: const <String>{}, cancelledServerIds: const <String>{});
+      return (
+        hubs: const <MediaHub>[],
+        observedItems: const <ObservedRow>[],
+        succeededServerIds: const <String>{},
+        cancelledServerIds: const <String>{},
+        failedServerIds: const <String>{},
+      );
     }
 
     // Home layout needs the library list for every client: fallback backends
@@ -444,9 +491,13 @@ class DataAggregationService {
     // (Plex) need it to detect visible music libraries, whose hubs the
     // global-hub endpoint excludes. One `fetchLibraries` per server, served
     // from the per-backend API cache when warm.
-    final libraries = useGlobalHubs
+    final libraryFetch = useGlobalHubs ? await getMediaLibrariesFromAllServers(serverIds: serverIds) : null;
-        ? _groupLibrariesByServer((await getMediaLibrariesFromAllServers(serverIds: serverIds)).libraries)
+    final libraries = libraryFetch == null ? null : _groupLibrariesByServer(libraryFetch.libraries);
-        : null;
+    final legSucceededServerIds = <String>{};
+    final legFailedServerIds = <String>{if (libraryFetch != null) ...libraryFetch.failedServerIds};
+    final legCancelledServerIds = <String>{if (libraryFetch != null) ...libraryFetch.cancelledServerIds};
+    final globalDiagnosticsByServer = <String, HubFetchDiagnostics>{};
+    final observedRows = <ObservedRow>[];
 
     final fetched = await _fanOut<MediaHub>(
       clients,
@@ -454,6 +505,7 @@ class DataAggregationService {
       fetch: (serverId, client) async {
         final serverLibraries = libraries?[serverId];
         final shouldUseGlobalHubs = useGlobalHubs && client.capabilities.richHubs;
+        final prefetchDegraded = legFailedServerIds.contains(serverId) || legCancelledServerIds.contains(serverId);
         final hubItemLimit = limit ?? defaultHubPreviewLimit;
         List<MediaHub> hubs;
         if (shouldUseGlobalHubs) {
@@ -461,12 +513,20 @@ class DataAggregationService {
           // Spreading `...await a, ...await b` into one list literal evaluates
           // them in order, which serialised the music rows behind the global
           // hub round trip.
-          final globalFuture = client.fetchGlobalHubs(limit: hubItemLimit, includePlaybackHubs: includePlaybackHubs);
+          final globalDiagnostics = HubFetchDiagnostics();
+          globalDiagnosticsByServer[serverId] = globalDiagnostics;
+          final globalFuture = client.fetchGlobalHubs(
+            limit: hubItemLimit,
+            includePlaybackHubs: includePlaybackHubs,
+            diagnostics: globalDiagnostics,
+          );
           // Plex's promoted/global hub endpoint never includes music
           // libraries — append their per-library hubs so music rows
-          // reach home. No-op (zero extra calls) without a visible
+          // reach home. A failed prefetch cannot establish that there
-          // music library.
+          // are no visible music libraries, so it is not a successful no-op.
-          final musicFuture = _fetchLibraryHubsForClient(
+          final musicFuture = prefetchDegraded
+              ? null
+              : _fetchLibraryHubsForClient(
                   client,
                   limit: hubItemLimit,
                   hiddenLibraryKeys: hiddenLibraryKeys,
@@ -474,23 +534,70 @@ class DataAggregationService {
                   libraries: serverLibraries ?? const [],
                   kinds: const {MediaKind.artist},
                 );
-          hubs = [...await globalFuture, ...await musicFuture];
+          Object? globalError;
+          StackTrace? globalStackTrace;
+          List<MediaHub> globalHubs = const [];
+          try {
+            globalHubs = await globalFuture;
+          } catch (error, stackTrace) {
+            globalDiagnostics.recordFailure(error);
+            globalError = error;
+            globalStackTrace = stackTrace;
+          }
+          final music = musicFuture == null ? null : await musicFuture;
+          if (globalHubs.isNotEmpty || (!globalDiagnostics.failed && !globalDiagnostics.cancelled)) {
+            legSucceededServerIds.add(serverId);
+          }
+          if (music != null) {
+            if (music.succeeded) legSucceededServerIds.add(serverId);
+            if (music.failed) legFailedServerIds.add(serverId);
+            if (music.cancelled) legCancelledServerIds.add(serverId);
+          }
+          if (globalError != null) Error.throwWithStackTrace(globalError, globalStackTrace!);
+          hubs = [...globalHubs, if (music != null) ...music.hubs];
         } else {
-          hubs = await _fetchLibraryHubsForClient(
+          // A fallback backend cannot run a hub leg without the libraries its
+          // prefetch failed to discover.
+          if (useGlobalHubs && prefetchDegraded) return const <MediaHub>[];
+          final libraryHubs = await _fetchLibraryHubsForClient(
             client,
             limit: hubItemLimit,
             hiddenLibraryKeys: hiddenLibraryKeys,
             includePlaybackHubs: includePlaybackHubs,
             libraries: useGlobalHubs ? serverLibraries : null,
           );
+          if (libraryHubs.succeeded || (!libraryHubs.failed && !libraryHubs.cancelled)) {
+            legSucceededServerIds.add(serverId);
           }
-        return _postProcessHubs(hubs, serverId: ServerId(serverId), hiddenLibraryKeys: hiddenLibraryKeys);
+          if (libraryHubs.failed) legFailedServerIds.add(serverId);
+          if (libraryHubs.cancelled) legCancelledServerIds.add(serverId);
+          hubs = libraryHubs.hubs;
+        }
+        final processed = _postProcessHubs(hubs, serverId: ServerId(serverId), hiddenLibraryKeys: hiddenLibraryKeys);
+        final scope = client.cacheServerId;
+        observedRows.addAll([
+          for (final hub in processed)
+            for (final item in hub.items) (item: item, clientScope: scope),
+        ]);
+        return processed;
       },
     );
 
+    for (final entry in globalDiagnosticsByServer.entries) {
+      if (entry.value.failed) legFailedServerIds.add(entry.key);
+      if (entry.value.cancelled) legCancelledServerIds.add(entry.key);
+    }
+
     final all = fetched.items;
     final hubs = limit != null && limit < all.length ? all.sublist(0, limit) : all;
-    return (hubs: hubs, succeededServerIds: fetched.succeededServerIds, cancelledServerIds: fetched.cancelledServerIds);
+    return (
+      hubs: hubs,
+      // Pre-limit, so a row trimmed off the tail still counts as observed.
+      observedItems: observedRows,
+      succeededServerIds: legSucceededServerIds,
+      cancelledServerIds: {...fetched.cancelledServerIds, ...legCancelledServerIds},
+      failedServerIds: {...fetched.failedServerIds, ...legFailedServerIds},
+    );
   }
 
   /// Per-library hub fetch for a single client. Filters to visible libraries
@@ -498,7 +605,12 @@ class DataAggregationService {
   /// musicvideos/homevideos, #1476; artist brings music rows to home) and
   /// concatenates the results. The rich-hub music append passes
   /// `{MediaKind.artist}` to fetch only what the global endpoint misses.
-  Future<List<MediaHub>> _fetchLibraryHubsForClient(
+  ///
+  /// `succeeded` stays false when no library leg was attempted. This prevents
+  /// the optional rich-hub music append from masking a failed global leg; the
+  /// fallback caller separately recognizes a clean zero-library result as an
+  /// authoritative no-op.
+  Future<({List<MediaHub> hubs, bool succeeded, bool failed, bool cancelled})> _fetchLibraryHubsForClient(
     MediaServerClient client, {
     required int limit,
     Set<String>? hiddenLibraryKeys,
@@ -524,29 +636,44 @@ class DataAggregationService {
     const concurrency = 3;
     final results = List<List<MediaHub>>.filled(visible.length, const []);
     var next = 0;
+    var succeeded = false;
+    var failed = false;
+    var cancelled = false;
 
     Future<void> worker() async {
       while (true) {
         final index = next++;
         if (index >= visible.length) return;
         final library = visible[index];
+        final diagnostics = HubFetchDiagnostics();
         try {
-          results[index] = await client.fetchLibraryHubs(
+          final hubs = await client.fetchLibraryHubs(
             library.id,
             libraryName: library.title,
             limit: limit,
             includePlaybackHubs: includePlaybackHubs,
             libraryKind: library.kind,
+            diagnostics: diagnostics,
           );
+          results[index] = hubs;
+          if (hubs.isNotEmpty || (!diagnostics.failed && !diagnostics.cancelled)) succeeded = true;
         } catch (e, st) {
+          if (_isCancellation(e)) {
+            cancelled = true;
+            appLogger.d('Cancelled library hub fetch for ${library.globalKey}');
+          } else {
+            failed = true;
             appLogger.e('Failed to fetch library hubs for ${library.globalKey}', error: e, stackTrace: st);
           }
         }
+        if (diagnostics.failed) failed = true;
+        if (diagnostics.cancelled) cancelled = true;
+      }
     }
 
     await Future.wait([for (var i = 0; i < concurrency && i < visible.length; i++) worker()]);
 
-    return [for (final list in results) ...list];
+    return (hubs: [for (final list in results) ...list], succeeded: succeeded, failed: failed, cancelled: cancelled);
   }
 
   /// Filter hidden-library items and drop empty hubs.

lib/services/external_player_service.dart

@@ -177,6 +177,7 @@ class ExternalPlayerService {
       return;
     }
 
+    var startedSucceeded = false;
     try {
       await client.reportPlaybackStarted(
         itemId: metadata.id,
@@ -185,6 +186,7 @@ class ExternalPlayerService {
         playMethod: 'DirectPlay',
         mediaSourceId: mediaSourceId,
       );
+      startedSucceeded = true;
     } catch (e) {
       appLogger.d('External player progress: started call failed (continuing)', error: e);
     }
@@ -210,6 +212,9 @@ class ExternalPlayerService {
       viewOffset: position.inMilliseconds,
       duration: duration.inMilliseconds,
       watchedThreshold: client.watchedThreshold,
+      // MediaBrowser persists stopped progress only for a session opened by
+      // Started; Plex persists every timeline report independently.
+      serverAcknowledged: !metadata.backend.usesMediaBrowserApi || startedSucceeded,
     );
 
     if (isWatchedProgress(

lib/services/jellyfin_client/parts/browse.dart

@@ -1593,7 +1593,11 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
   }
 
   @override
-  Future<List<MediaHub>> fetchGlobalHubs({int limit = defaultHubPreviewLimit, bool includePlaybackHubs = true}) async {
+  Future<List<MediaHub>> fetchGlobalHubs({
+    int limit = defaultHubPreviewLimit,
+    bool includePlaybackHubs = true,
+    HubFetchDiagnostics? diagnostics,
+  }) async {
     // Jellyfin doesn't expose a single "hubs" endpoint, so we synthesise the
     // home rows from Latest plus optional playback rows. The richer Plex Discover surface
     // is intentionally left untranslated — see ServerCapabilities.richHubs.
@@ -1607,6 +1611,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
       continueTitle: t.discover.continueWatching,
       nextUpTitle: t.discover.nextUp,
       recentTitle: t.discover.recentlyAdded,
+      diagnostics: diagnostics,
     );
   }
 
@@ -1617,6 +1622,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     int limit = defaultHubPreviewLimit,
     bool includePlaybackHubs = true,
     MediaKind? libraryKind,
+    HubFetchDiagnostics? diagnostics,
   }) async {
     // Music libraries get their own hub set. Home passes
     // includePlaybackHubs=false because it already renders the app-level
@@ -1630,6 +1636,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
         libraryName: libraryName,
         limit: limit,
         includePlaybackHubs: includePlaybackHubs,
+        diagnostics: diagnostics,
       );
     }
 
@@ -1649,6 +1656,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
       continueTitle: t.discover.continueWatchingIn(library: libraryName),
       nextUpTitle: t.discover.nextUpIn(library: libraryName),
       recentTitle: t.discover.recentlyAddedIn(library: libraryName),
+      diagnostics: diagnostics,
     );
   }
 
@@ -1670,14 +1678,20 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     required String recentTitle,
     String? parentId,
     String? latestItemTypes,
+    HubFetchDiagnostics? diagnostics,
   }) async {
-    final latestFuture = _safeFetchItemsArray('/Users/${_segment(connection.userId)}/Items/Latest', {
+    final latestFuture = _safeFetchItemsArray(
+      '/Users/${_segment(connection.userId)}/Items/Latest',
+      {
         'Limit': limit.toString(),
         'ParentId': ?parentId,
         'Fields': _hubRowFields,
         'IncludeItemTypes': ?latestItemTypes,
         ...jellyfinImageQueryParameters,
-    }, retry: retry);
+      },
+      retry: retry,
+      diagnostics: diagnostics,
+    );
 
     MediaHub hub(String suffix, String title, String type, List<Map<String, dynamic>> items) =>
         JellyfinMappers.syntheticHub(
@@ -1698,7 +1712,9 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
 
     final results = await Future.wait([
       latestFuture,
-      _safeFetchItemsArray(_resumePath, {
+      _safeFetchItemsArray(
+        _resumePath,
+        {
           'userId': connection.userId,
           ..._resumeFilterQuery,
           'ParentId': ?parentId,
@@ -1708,9 +1724,13 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
           'Recursive': 'true',
           'EnableTotalRecordCount': 'false',
           ...jellyfinImageQueryParameters,
-      }, retry: retry),
+        },
+        retry: retry,
+        diagnostics: diagnostics,
+      ),
       includeNextUp
-          ? _fetchNextUpRows({
+          ? _fetchNextUpRows(
+              {
                 'userId': connection.userId,
                 'ParentId': ?parentId,
                 'Limit': limit.toString(),
@@ -1719,7 +1739,10 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
                 'NextUpDateCutoff': _nextUpDateCutoff(),
                 'EnableTotalRecordCount': 'false',
                 ...jellyfinImageQueryParameters,
-            }, retry: retry)
+              },
+              retry: retry,
+              diagnostics: diagnostics,
+            )
           : Future.value(const <Map<String, dynamic>>[]),
     ]);
 
@@ -1744,14 +1767,20 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     required String libraryName,
     required int limit,
     required bool includePlaybackHubs,
+    HubFetchDiagnostics? diagnostics,
   }) async {
-    final latestFuture = _safeFetchItemsArray('/Users/${_segment(connection.userId)}/Items/Latest', {
+    final latestFuture = _safeFetchItemsArray(
+      '/Users/${_segment(connection.userId)}/Items/Latest',
+      {
         'Limit': limit.toString(),
         'ParentId': libraryId,
         'Fields': _musicAlbumRowFields,
         'EnableUserData': 'false',
         ...jellyfinImageQueryParameters,
-    }, retry: _libraryHubRetry);
+      },
+      retry: _libraryHubRetry,
+      diagnostics: diagnostics,
+    );
 
     MediaHub latestAlbumsHub(List<Map<String, dynamic>> items) => JellyfinMappers.syntheticHub(
       mapItem: _mapItem,
@@ -1781,8 +1810,18 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     };
     final results = await Future.wait([
       latestFuture,
-      _safeFetchItemsArray('/Items', {...playedParams, 'SortBy': 'DatePlayed'}, retry: _libraryHubRetry),
+      _safeFetchItemsArray(
-      _safeFetchItemsArray('/Items', {...playedParams, 'SortBy': 'PlayCount'}, retry: _libraryHubRetry),
+        '/Items',
+        {...playedParams, 'SortBy': 'DatePlayed'},
+        retry: _libraryHubRetry,
+        diagnostics: diagnostics,
+      ),
+      _safeFetchItemsArray(
+        '/Items',
+        {...playedParams, 'SortBy': 'PlayCount'},
+        retry: _libraryHubRetry,
+        diagnostics: diagnostics,
+      ),
     ]);
 
     return [
@@ -2155,9 +2194,16 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     Map<String, dynamic> queryParameters, {
     _HubRetryPolicy? retry,
     AbortController? abort,
+    HubFetchDiagnostics? diagnostics,
   }) async {
     if (dialect.supportsGlobalNextUp) {
-      return _safeFetchItemsArray('/Shows/NextUp', queryParameters, retry: retry, abort: abort);
+      return _safeFetchItemsArray(
+        '/Shows/NextUp',
+        queryParameters,
+        retry: retry,
+        abort: abort,
+        diagnostics: diagnostics,
+      );
     }
 
     final budgetAbort = AbortController();
@@ -2508,6 +2554,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
     Map<String, dynamic> queryParameters, {
     _HubRetryPolicy? retry,
     AbortController? abort,
+    HubFetchDiagnostics? diagnostics,
     Duration? timeout,
     bool allowEndpointFailover = true,
   }) async {
@@ -2532,6 +2579,7 @@ mixin _JellyfinBrowseMethods on _JellyfinClientInternals {
       // it propagate so the caller classifies the fetch as disrupted, not
       // empty.
       if (e is MediaServerHttpException && e.isCancellation) rethrow;
+      diagnostics?.recordFailure(e);
       appLogger.w('JellyfinClient: $path failed (treating as empty)', error: e, stackTrace: st);
       return const [];
     }

lib/services/jellyfin_client/parts/live_tv.dart

@@ -12,6 +12,8 @@ mixin _JellyfinLiveTvMethods on _JellyfinClientInternals {
     Duration? timeout,
     // ignore: unused_element_parameter
     bool allowEndpointFailover,
+    // ignore: unused_element_parameter
+    HubFetchDiagnostics? diagnostics,
   });
 
   /// Returns `true` when this server has Live TV configured (channels

lib/services/offline_watch_sync_service.dart

@@ -22,6 +22,17 @@ import 'settings_service.dart';
 import 'trackers/tracker_coordinator.dart';
 import 'watch_state_resolver.dart';
 
+typedef QueuedOfflineWatchAction = ({String? clientScopeId, String? profileId, int rowId, int revision});
+
+typedef _OfflineWatchReplayResult = ({
+  MediaItem item,
+  String? clientScopeId,
+  String? profileId,
+  int rowId,
+  int revision,
+  bool persisted,
+});
+
 /// Service for managing offline watch progress and syncing it back to the
 /// owning server. Backend-neutral over [MediaServerClient] — Plex actions
 /// hit `/:/scrobble` and `/:/timeline`, while MediaBrowser actions use their
@@ -110,16 +121,30 @@ class OfflineWatchSyncService extends ChangeNotifier {
     if (event.changeType != WatchStateChangeType.watched && event.changeType != WatchStateChangeType.unwatched) {
       return;
     }
-    unawaited(_discardQueuedProgress(ServerId(event.serverId), event.itemId));
+    unawaited(
+      _discardQueuedProgress(
+        ServerId(event.serverId),
+        event.itemId,
+        beforeRevision: _offlineActionRevision(event.patchId),
+      ),
+    );
   }
 
-  Future<void> _discardQueuedProgress(ServerId serverId, String itemId) async {
+  int? _offlineActionRevision(WatchPatchId? patchId) {
+    final value = patchId?.value;
+    if (value == null || !value.startsWith('o:')) return null;
+    final separator = value.lastIndexOf(':');
+    return separator < 2 ? null : int.tryParse(value.substring(separator + 1));
+  }
+
+  Future<void> _discardQueuedProgress(ServerId serverId, String itemId, {int? beforeRevision}) async {
     try {
       final removed = await _database.deleteQueuedProgressForItem(
         profileId: _activeProfileId,
         serverId: serverId,
         clientScopeId: await _clientScopeIdForItem(serverId, itemId),
         ratingKey: itemId,
+        beforeRevision: beforeRevision,
       );
       if (removed == 0) return;
       appLogger.d('Dropped $removed superseded queued progress action(s) for $serverId:$itemId');
@@ -258,7 +283,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
   /// The `ratingKey:` parameter on the underlying `_database` calls is
   /// preserved as the on-disk column name; the in-memory parameter renamed
   /// here is just the API-level identifier.
-  Future<String?> queueProgressUpdate({
+  Future<QueuedOfflineWatchAction> queueProgressUpdate({
     required ServerId serverId,
     required String itemId,
     required int viewOffset,
@@ -267,9 +292,10 @@ class OfflineWatchSyncService extends ChangeNotifier {
     final shouldMarkWatched =
         duration != null && isWatchedByProgress(viewOffset, duration, serverId: ServerId(serverId));
     final clientScopeId = await _clientScopeIdForItem(ServerId(serverId), itemId);
+    final profileId = _activeProfileId;
 
-    await _database.upsertProgressAction(
+    final queued = await _database.upsertProgressAction(
-      profileId: _activeProfileId,
+      profileId: profileId,
       serverId: serverId,
       clientScopeId: clientScopeId,
       ratingKey: itemId,
@@ -287,23 +313,24 @@ class OfflineWatchSyncService extends ChangeNotifier {
     );
 
     notifyListeners();
-    return clientScopeId;
+    return (clientScopeId: clientScopeId, profileId: profileId, rowId: queued.rowId, revision: queued.revision);
   }
 
-  Future<String?> queueMarkWatched({required ServerId serverId, required String itemId}) =>
+  Future<QueuedOfflineWatchAction> queueMarkWatched({required ServerId serverId, required String itemId}) =>
       _queueWatchStatusAction(serverId: serverId, itemId: itemId, actionType: OfflineActionType.watched.id);
 
-  Future<String?> queueMarkUnwatched({required ServerId serverId, required String itemId}) =>
+  Future<QueuedOfflineWatchAction> queueMarkUnwatched({required ServerId serverId, required String itemId}) =>
       _queueWatchStatusAction(serverId: serverId, itemId: itemId, actionType: OfflineActionType.unwatched.id);
 
-  Future<String?> _queueWatchStatusAction({
+  Future<QueuedOfflineWatchAction> _queueWatchStatusAction({
     required ServerId serverId,
     required String itemId,
     required String actionType,
   }) async {
     final clientScopeId = await _clientScopeIdForItem(ServerId(serverId), itemId);
-    await _database.insertWatchAction(
+    final profileId = _activeProfileId;
-      profileId: _activeProfileId,
+    final queued = await _database.insertWatchAction(
+      profileId: profileId,
       serverId: serverId,
       clientScopeId: clientScopeId,
       ratingKey: itemId,
@@ -312,7 +339,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
 
     appLogger.d('Queued offline mark $actionType: $serverId:$itemId');
     notifyListeners();
-    return clientScopeId;
+    return (clientScopeId: clientScopeId, profileId: profileId, rowId: queued.rowId, revision: queued.revision);
   }
 
   /// Check if an item should be considered watched based on progress percentage.
@@ -446,14 +473,29 @@ class OfflineWatchSyncService extends ChangeNotifier {
           continue;
         }
 
-        final synced = await _withOnlineClientForAction(action, (client) async {
+        final synced = await _withOnlineClientForAction(action, (client, clientScopeId) async {
           try {
-            await _syncAction(client, action);
+            final result = await _syncAction(client, action, clientScopeId: clientScopeId);
-            await _database.deleteWatchAction(action.id);
+            if (!result.persisted) {
+              // The write did not land — MediaBrowser drops a stop for a
+              // session its Started never opened. Leave the row queued and
+              // count the attempt so a later pass retries it.
+              appLogger.d('Action ${action.id} did not persist; keeping it queued');
+              await _database.updateSyncAttemptIfUnchanged(action.id, action.updatedAt, 'write did not persist');
+              return;
+            }
+            final deleted = await _database.deleteWatchActionIfUnchanged(result.rowId, result.revision);
+            if (!deleted) {
+              appLogger.d('Synced action ${action.id} revision ${action.updatedAt}; a newer revision remains queued');
+              return;
+            }
+            WatchPatchPromotionNotifier().promote(
+              WatchPatchId.offlineAction(profileId: result.profileId, rowId: result.rowId, revision: result.revision),
+            );
             appLogger.d('Successfully synced action ${action.id}: ${action.actionType} for ${action.ratingKey}');
           } catch (e) {
             appLogger.w('Failed to sync action ${action.id}: $e');
-            await _database.updateSyncAttempt(action.id, e.toString());
+            await _database.updateSyncAttemptIfUnchanged(action.id, action.updatedAt, e.toString());
           }
         });
         if (!synced) {
@@ -507,7 +549,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
 
   Future<bool> _withOnlineClientForAction(
     OfflineWatchProgressItem action,
-    Future<void> Function(MediaServerClient client) callback,
+    Future<void> Function(MediaServerClient client, String? clientScopeId) callback,
   ) async {
     final resolved = await _clientForAction(action);
     if (resolved == null) {
@@ -520,7 +562,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
       return false;
     }
 
-    await callback(resolved.client);
+    await callback(resolved.client, resolved.clientScopeId);
     return true;
   }
 
@@ -575,14 +617,14 @@ class OfflineWatchSyncService extends ChangeNotifier {
   /// Uses the neutral [MediaServerClient] surface so Jellyfin's
   /// `/UserPlayedItems/{id}` and `/Sessions/Playing*` endpoints receive
   /// the same queued state Plex's `/:/scrobble` and `/:/timeline` do.
-  Future<void> _syncAction(MediaServerClient client, OfflineWatchProgressItem action) async {
+  Future<_OfflineWatchReplayResult> _syncAction(
-    // Fetch metadata so trackers (and the stop-path watch event) get enough
+    MediaServerClient client,
-    // context — external ids, parent chain, library section. The plain
+    OfflineWatchProgressItem action, {
-    // watched/unwatched replays deliberately emit no WatchStateEvent: the
+    required String? clientScopeId,
-    // offline provider already emitted it when the action was queued, and
+  }) async {
-    // client markWatched/markUnwatched are transport-only. Best-effort: a
+    // Fetch metadata so tracker writes get enough context — external ids,
-    // missed metadata fetch falls back to a minimal MediaItem — the network
+    // parent chain and library section. Best-effort: a missed metadata fetch
-    // call still goes through.
+    // falls back to a minimal item while the server write still proceeds.
     final needsRichItem =
         action.actionType == OfflineActionType.watched.id ||
         action.actionType == OfflineActionType.unwatched.id ||
@@ -602,14 +644,17 @@ class OfflineWatchSyncService extends ChangeNotifier {
       serverId: action.serverId,
     );
 
+    var persisted = false;
     switch (action.actionType) {
       case 'watched':
         await client.markWatched(item);
+        persisted = true;
         await TrackerCoordinator.instance.markWatched(item, client);
         break;
 
       case 'unwatched':
         await client.markUnwatched(item);
+        persisted = true;
         await TrackerCoordinator.instance.markUnwatched(item, client);
         break;
 
@@ -622,13 +667,19 @@ class OfflineWatchSyncService extends ChangeNotifier {
           final position = action.shouldMarkWatched && duration != null
               ? duration
               : Duration(milliseconds: action.viewOffset!);
-          if (!action.shouldMarkWatched || client.backend.usesMediaBrowserApi) {
+          // MediaBrowser ignores a stop for a session it never opened, so its
+          // Started call is a precondition for persistence, not best effort.
+          final requiresOpenSession = client.backend.usesMediaBrowserApi;
+          var startedSucceeded = !requiresOpenSession;
+          if (!action.shouldMarkWatched || requiresOpenSession) {
             try {
               await client.reportPlaybackStarted(itemId: action.ratingKey, position: position, duration: duration);
+              startedSucceeded = true;
             } catch (e) {
-              // Plex sometimes 5xxs the start when nothing follows; treat as
+              // Plex sometimes 5xxs the start when nothing follows; there the
-              // best-effort and continue to the stop call which is the one
+              // stop call is the one that persists the resume position, so
-              // that actually persists the resume position.
+              // continue. On MediaBrowser the stop will be dropped, so the
+              // action must stay queued for a later attempt.
               appLogger.d('Offline progress: started call failed (continuing)', error: e);
             }
           }
@@ -640,18 +691,30 @@ class OfflineWatchSyncService extends ChangeNotifier {
               recordedAt: DateTime.fromMillisecondsSinceEpoch(action.updatedAt),
             ),
           );
+          persisted = startedSucceeded;
         }
 
-        // If progress exceeded threshold, also mark as watched. On backends
-        // that mark played from the stopped report above (MediaBrowser), this
-        // only emits the local watch event — an explicit markWatched would
-        // double-scrobble via the Trakt plugin (#1287).
         if (action.shouldMarkWatched) {
-          await client.markWatchedFromPlaybackStop(item);
+          // MediaBrowser persists the played state from the stopped report.
+          // Plex still needs the explicit transport call, but replay must not
+          // emit another semantic event for an action already emitted offline.
+          if (!client.marksWatchedOnPlaybackStopped) {
+            await client.markWatched(item);
+            persisted = true;
+          }
           await TrackerCoordinator.instance.markWatched(item, client);
         }
         break;
     }
+
+    return (
+      item: item,
+      clientScopeId: clientScopeId,
+      profileId: action.profileId,
+      rowId: action.id,
+      revision: action.updatedAt,
+      persisted: persisted,
+    );
   }
 
   /// Push a watch-state change Plezy observed on the server to the trackers.
@@ -703,6 +766,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
               item: episode,
               isNowWatched: isWatched,
               cacheServerId: client.cacheServerId,
+              serverAcknowledged: true,
             );
             // The change came from the server (watched on another client), so no
             // playback or manual path has told the trackers about it.
@@ -841,6 +905,7 @@ class OfflineWatchSyncService extends ChangeNotifier {
                     item: metadata,
                     isNowWatched: isWatched,
                     cacheServerId: client.cacheServerId,
+                    serverAcknowledged: true,
                   );
                   await _mirrorWatchStateToTrackers(metadata, client, isWatched: isWatched);
                 }

lib/services/playback_progress_tracker.dart

@@ -146,6 +146,24 @@ class PlaybackProgressTracker {
   /// The post-watch hook has run; it fires at most once per tracker.
   bool _scrobbledHookRan = false;
 
+  /// Delivery provenance keyed by the exact snapshot handed to
+  /// [PlaybackReportSession]. Acceptance alone cannot identify delivery
+  /// because startup heartbeats may be dropped while still completing true.
+  final Map<PlaybackReportSnapshot, bool> _deliveredProgressAcknowledgements = Map.identity();
+
+  /// Whether this backend reporting session has successfully opened.
+  bool _hasDeliveredStart = false;
+
+  /// Whether the delivered stopped report persisted its position.
+  bool _stoppedProgressServerAcknowledged = false;
+
+  /// Allows the first persisted Progress after a MediaBrowser Started to
+  /// upgrade local provenance even when the position delta is throttled.
+  bool _lastProgressNotificationServerAcknowledged = false;
+
+  /// The exact report-derived watched patch that an explicit mark can settle.
+  WatchPatchId? _watchedPatchId;
+
   /// Whether the final stopped progress event was already emitted locally.
   bool _stopProgressNotified = false;
 
@@ -261,6 +279,8 @@ class PlaybackProgressTracker {
     _deliveredBelow = false;
     _serverObservedCrossing = false;
     _sessionEnded = false;
+    _hasDeliveredStart = false;
+    _stoppedProgressServerAcknowledged = false;
   }
 
   Future<void> _sendProgress(String state, {Duration? positionOverride}) async {
@@ -292,7 +312,7 @@ class PlaybackProgressTracker {
         // nothing.
         if (!canCommitStoppedProgress) return;
         await _sendOfflineProgress(position, duration);
-        _notifyProgressIfNeeded(position, duration, force: state == 'stopped');
+        _notifyProgressIfNeeded(position, duration, force: state == 'stopped', serverAcknowledged: false);
       } else if (state == 'stopped') {
         // Stopped must complete before disposal. When reporting was disabled
         // by a fatal error, use the last position captured while output was
@@ -304,17 +324,19 @@ class PlaybackProgressTracker {
         await _pendingSettle;
         _resetBackoff();
         if (accepted && canCommitStoppedProgress) {
-          _notifyProgressIfNeeded(position, duration, force: true);
+          _notifyProgressIfNeeded(
+            position,
+            duration,
+            force: true,
+            serverAcknowledged: _stoppedProgressServerAcknowledged,
+          );
         }
       } else {
         // Fire-and-forget for playing/paused — avoid blocking the Dart event loop
         unawaited(
           _sendOnlineProgress(state, position, duration)
-              .then((accepted) {
+              .then((_) {
                 _resetBackoff();
-                if (accepted) {
-                  _notifyProgressIfNeeded(position, duration);
-                }
               })
               .catchError((Object e) {
                 _recordProgressFailure(e);
@@ -370,7 +392,12 @@ class PlaybackProgressTracker {
     }
   }
 
-  void _notifyProgressIfNeeded(Duration position, Duration duration, {bool force = false}) {
+  void _notifyProgressIfNeeded(
+    Duration position,
+    Duration duration, {
+    bool force = false,
+    required bool serverAcknowledged,
+  }) {
     if (_scrobbled) return;
     if (position.inMilliseconds <= 0 || duration.inMilliseconds <= 0) return;
     if (force) {
@@ -378,16 +405,20 @@ class PlaybackProgressTracker {
       _stopProgressNotified = true;
     } else {
       final last = _lastProgressNotifiedPosition;
-      if (last != null && (position - last).abs() < _progressNotifyDelta) return;
+      if (last != null && (position - last).abs() < _progressNotifyDelta) {
+        if (!serverAcknowledged || _lastProgressNotificationServerAcknowledged) return;
+      }
     }
 
     _lastProgressNotifiedPosition = position;
+    _lastProgressNotificationServerAcknowledged = serverAcknowledged;
     WatchStateNotifier().notifyProgress(
       item: metadata,
       cacheServerId: client?.cacheServerId,
       viewOffset: position.inMilliseconds,
       duration: duration.inMilliseconds,
       watchedThreshold: client?.watchedThreshold ?? 0.9,
+      serverAcknowledged: serverAcknowledged,
     );
   }
 
@@ -403,20 +434,26 @@ class PlaybackProgressTracker {
     final session = _reportSession;
     if (c == null || session == null) return false;
 
-    final accepted = await session.report(
+    final snapshot = PlaybackReportSnapshot(
-      PlaybackReportSnapshot(
       state: state,
       position: position,
       duration: duration,
       resolveStreamSelection: state == 'stopped'
           ? _currentStreamSelectionForStopped
           : _currentStreamSelectionForProgress,
-      ),
     );
+    final accepted = await session.report(snapshot);
 
     if (accepted && allowScrobble) {
       await _maybeScrobble(c, position, duration);
     }
+
+    if (!snapshot.isStopped) {
+      final serverAcknowledged = _deliveredProgressAcknowledgements.remove(snapshot);
+      if (serverAcknowledged != null) {
+        _notifyProgressIfNeeded(position, duration, serverAcknowledged: serverAcknowledged);
+      }
+    }
     return accepted;
   }
 
@@ -434,7 +471,24 @@ class PlaybackProgressTracker {
   /// whose every report sits above the threshold, or one resuming past it, is
   /// never marked server-side.
   void _onReportDelivered(PlaybackReportSnapshot snapshot) {
-    final threshold = client?.watchedThreshold;
+    final c = client;
+    // The same backend as the client's, but read from the item so the
+    // classification matches the pattern already used for track selection
+    // below and does not depend on a client method.
+    final persistsPositionOnEveryReport = !metadata.backend.usesMediaBrowserApi;
+    if (snapshot.isStopped) {
+      // MediaBrowser needs a successfully opened session before Stopped can
+      // persist position; Plex timeline reports are independent.
+      _stoppedProgressServerAcknowledged = persistsPositionOnEveryReport || _hasDeliveredStart;
+    } else {
+      final isStarted = !_hasDeliveredStart;
+      _hasDeliveredStart = true;
+      // A MediaBrowser `Started` saves play count and last-played date but
+      // deliberately not the position, so it cannot acknowledge an offset.
+      _deliveredProgressAcknowledgements[snapshot] = !isStarted || persistsPositionOnEveryReport;
+    }
+
+    final threshold = c?.watchedThreshold;
     // isWatchedProgress reports false for an unknown duration; treating that as
     // a below-threshold report would wrongly arm the crossing.
     if (threshold == null || snapshot.duration.inMilliseconds <= 0) return;
@@ -503,6 +557,8 @@ class PlaybackProgressTracker {
       _serverMarkSettled = false; // Retry on the next delivered report.
       return;
     }
+    final patchId = _watchedPatchId;
+    if (patchId != null) WatchPatchPromotionNotifier().promote(patchId);
     await _runScrobbledHook();
   }
 
@@ -540,7 +596,7 @@ class PlaybackProgressTracker {
     // Local state flips on the observed crossing, whether or not the backend
     // received that particular report. The server-side mark is a separate
     // question, answered by _settleServerMark once delivery is known.
-    c.notifyWatchedFromPlaybackSession(metadata);
+    _watchedPatchId = c.notifyWatchedFromPlaybackSession(metadata);
     appLogger.d(
       'Watched ${metadata.id} (${(percent * 100).toStringAsFixed(0)}% >= ${(threshold * 100).toStringAsFixed(0)}%)',
     );

lib/services/plex_client.dart

@@ -2226,6 +2226,7 @@ class PlexClient
     int? librarySectionID,
     String? librarySectionTitle,
     bool Function(PlexMetadataDto)? filter,
+    HubFetchDiagnostics? diagnostics,
   }) async {
     try {
       final response = await retryTransientMediaServerCall(
@@ -2253,6 +2254,7 @@ class PlexClient
         ),
       );
     } catch (e) {
+      diagnostics?.recordFailure(e);
       appLogger.e('Failed to get $failureLabel: $e');
     }
     return [];
@@ -2264,6 +2266,7 @@ class PlexClient
     String sectionId, {
     int limit = defaultHubPreviewLimit,
     String? libraryName,
+    HubFetchDiagnostics? diagnostics,
   }) => _fetchHubs(
     path: '/hubs/sections/$sectionId',
     queryParameters: {'count': limit, 'includeGuids': 1},
@@ -2272,18 +2275,21 @@ class PlexClient
     failureLabel: 'library hubs',
     librarySectionID: _librarySectionIdFromString(sectionId),
     librarySectionTitle: libraryName,
+    diagnostics: diagnostics,
     filter: _videoOrMusicHubItem,
   );
 
   /// Get global hubs (home page recommendations)
   /// Returns actual home page hubs like "Recently Added Movies", "Recently Added TV", etc.
   /// This matches the official Plex client's home page layout.
-  Future<List<PlexHubDto>> _getGlobalHubs({int limit = defaultHubPreviewLimit}) => _fetchHubs(
+  Future<List<PlexHubDto>> _getGlobalHubs({int limit = defaultHubPreviewLimit, HubFetchDiagnostics? diagnostics}) =>
+      _fetchHubs(
         path: _providerPromotedHubKey ?? _providerHomeHubKey ?? '/hubs',
         queryParameters: {'count': limit, 'includeGuids': 1},
         operation: 'Plex global hubs',
         deadline: MediaServerTimeouts.homeHubDeadline,
         failureLabel: 'global hubs',
+        diagnostics: diagnostics,
       );
 
   /// Get related hubs for a specific metadata item (collections, similar, "more from" director/actor)
@@ -3614,8 +3620,12 @@ class PlexClient
   }
 
   @override
-  Future<List<MediaHub>> fetchGlobalHubs({int limit = defaultHubPreviewLimit, bool includePlaybackHubs = true}) async {
+  Future<List<MediaHub>> fetchGlobalHubs({
-    final hubs = await _getGlobalHubs(limit: limit);
+    int limit = defaultHubPreviewLimit,
+    bool includePlaybackHubs = true,
+    HubFetchDiagnostics? diagnostics,
+  }) async {
+    final hubs = await _getGlobalHubs(limit: limit, diagnostics: diagnostics);
     return hubs.map((h) => PlexMappers.mediaHub(h)).toList();
   }
 
@@ -3626,10 +3636,11 @@ class PlexClient
     int limit = defaultHubPreviewLimit,
     bool includePlaybackHubs = true,
     MediaKind? libraryKind,
+    HubFetchDiagnostics? diagnostics,
   }) async {
     // libraryName is unused: Plex's /hubs/sections/{id} returns hubs already
     // titled per-library (e.g. "Recently Added in Movies").
-    final hubs = await _getLibraryHubs(libraryId, limit: limit, libraryName: libraryName);
+    final hubs = await _getLibraryHubs(libraryId, limit: limit, libraryName: libraryName, diagnostics: diagnostics);
     return hubs.map((h) => PlexMappers.mediaHub(h)).toList();
   }
 

lib/services/watch_actions.dart

@@ -63,7 +63,12 @@ class WatchActions {
     } else {
       await client.markUnwatched(item);
     }
-    WatchStateNotifier().notifyWatched(item: item, isNowWatched: watched, cacheServerId: client.cacheServerId);
+    WatchStateNotifier().notifyWatched(
+      item: item,
+      isNowWatched: watched,
+      cacheServerId: client.cacheServerId,
+      serverAcknowledged: true,
+    );
     unawaited(
       watched
           ? TrackerCoordinator.instance.markWatched(item, client)

lib/utils/watch_state_notifier.dart

@@ -7,6 +7,39 @@ import 'global_key_utils.dart';
 import 'hierarchical_event_mixin.dart';
 import 'media_event_keys.dart';
 
+/// Identity of the store overlay entry a watch event creates.
+///
+/// Two forms, because the two paths that need to settle later have different
+/// lifetimes:
+/// - [WatchPatchId.session] is minted per event and dies with the process,
+///   which is right for a live playback crossing.
+/// - [WatchPatchId.offlineAction] is *derived* from the persisted queue row
+///   `(profileId, rowId, revision)` rather than minted, so emission,
+///   hydration and promotion still join after a restart. A minted id could
+///   not: the database persists the row and its revision, nothing more.
+///
+/// Promotion matches on this, never on the global key — a same-item rewatch
+/// must never be promoted or discarded in place of the action that settled.
+class WatchPatchId {
+  final String value;
+
+  const WatchPatchId._(this.value);
+
+  factory WatchPatchId.session(int sequence) => WatchPatchId._('s:$sequence');
+
+  factory WatchPatchId.offlineAction({required String? profileId, required int rowId, required int revision}) =>
+      WatchPatchId._('o:${profileId ?? ''}:$rowId:$revision');
+
+  @override
+  bool operator ==(Object other) => other is WatchPatchId && other.value == value;
+
+  @override
+  int get hashCode => value.hashCode;
+
+  @override
+  String toString() => 'WatchPatchId($value)';
+}
+
 enum WatchStateChangeType { watched, unwatched, progressUpdate, removedFromContinueWatching }
 
 /// Event representing a watch state change with parent chain for hierarchical invalidation
@@ -53,6 +86,20 @@ class WatchStateEvent with HierarchicalEventMixin {
   /// numeric id, Jellyfin sends a UUID; both round-trip as strings.
   final String? librarySectionID;
 
+  /// Whether the server had already accepted this exact state when the event
+  /// was emitted.
+  ///
+  /// Only an acknowledged patch may be superseded by a later authoritative
+  /// read (#1829): an unacknowledged one represents a write still owed to the
+  /// server, and a read must not retire it. Defaults to `false` so a new emit
+  /// site that forgets to classify itself degrades to today's behaviour rather
+  /// than silently becoming retireable.
+  final bool serverAcknowledged;
+
+  /// Identity of the overlay entry this event creates, for later promotion.
+  /// Null for events nothing will ever settle.
+  final WatchPatchId? patchId;
+
   WatchStateEvent({
     required this.itemId,
     required this.serverId,
@@ -63,6 +110,8 @@ class WatchStateEvent with HierarchicalEventMixin {
     this.viewOffset,
     this.isNowWatched,
     this.librarySectionID,
+    this.serverAcknowledged = false,
+    this.patchId,
   }) : globalKey = buildGlobalKey(ServerId(serverId), itemId);
 
   /// `serverId:librarySectionID`, matching [MediaLibrary.globalKey]. Null when
@@ -98,9 +147,20 @@ class WatchStateNotifier extends BaseNotifier<WatchStateEvent> {
   }
 
   /// Helper to emit a watched/unwatched event from a [MediaItem].
-  void notifyWatched({required MediaItem item, bool isNowWatched = true, String? cacheServerId}) {
+  ///
+  /// Returns the [WatchPatchId] of the overlay entry it created so a caller
+  /// that must settle later can promote that exact entry. Callers with
+  /// nothing to settle ignore it.
+  WatchPatchId? notifyWatched({
+    required MediaItem item,
+    bool isNowWatched = true,
+    String? cacheServerId,
+    bool serverAcknowledged = false,
+    WatchPatchId? patchId,
+  }) {
     final serverId = serverIdForEvent(item, notifier: 'WatchStateNotifier', event: 'watched');
-    if (serverId == null) return;
+    if (serverId == null) return null;
+    final id = patchId ?? WatchPatchId.session(++_sequence);
     notify(
       WatchStateEvent(
         itemId: item.id,
@@ -111,23 +171,34 @@ class WatchStateNotifier extends BaseNotifier<WatchStateEvent> {
         mediaType: item.kind.id,
         isNowWatched: isNowWatched,
         librarySectionID: item.libraryId,
+        serverAcknowledged: serverAcknowledged,
+        patchId: id,
       ),
     );
+    return id;
   }
 
+  /// Monotonic source for session-minted [WatchPatchId]s.
+  var _sequence = 0;
+
   /// Helper to emit a progress update event.
   /// [watchedThreshold] defaults to 0.9 — pass the server's configured value
   /// (`client.watchedThreshold`) when available.
-  void notifyProgress({
+  ///
+  /// Returns the created [WatchPatchId], as [notifyWatched] does.
+  WatchPatchId? notifyProgress({
     required MediaItem item,
     required int viewOffset,
     required int duration,
     String? cacheServerId,
     double watchedThreshold = 0.9,
+    bool serverAcknowledged = false,
+    WatchPatchId? patchId,
   }) {
     final serverId = serverIdForEvent(item, notifier: 'WatchStateNotifier', event: 'progress');
-    if (serverId == null) return;
+    if (serverId == null) return null;
     final isNowWatched = isWatchedProgress(positionMs: viewOffset, durationMs: duration, threshold: watchedThreshold);
+    final id = patchId ?? WatchPatchId.session(++_sequence);
 
     notify(
       WatchStateEvent(
@@ -140,8 +211,11 @@ class WatchStateNotifier extends BaseNotifier<WatchStateEvent> {
         viewOffset: viewOffset,
         isNowWatched: isNowWatched,
         librarySectionID: item.libraryId,
+        serverAcknowledged: serverAcknowledged,
+        patchId: id,
       ),
     );
+    return id;
   }
 
   /// Helper to emit a Continue Watching removal event.
@@ -160,3 +234,37 @@ class WatchStateNotifier extends BaseNotifier<WatchStateEvent> {
     );
   }
 }
+
+/// A patch the server has now definitively accepted.
+///
+/// Deliberately *not* a [WatchStateEvent]: semantic subscribers must not see
+/// promotions. `OfflineWatchSyncService` reacts to `watched`/`unwatched` by
+/// purging queued progress, so replaying one here would delete a newer
+/// rewatch — the very write promotion exists to protect.
+///
+/// [BaseNotifier] is single-stream by construction, and `WatchStateNotifier`
+/// fixes its type to [WatchStateEvent], so this needs its own channel rather
+/// than a discriminated union on the existing one.
+class WatchPatchPromotion {
+  final WatchPatchId patchId;
+
+  const WatchPatchPromotion(this.patchId);
+
+  @override
+  String toString() => 'WatchPatchPromotion($patchId)';
+}
+
+/// Sibling singleton to [WatchStateNotifier] carrying non-semantic
+/// promotions, following the same pattern as `LibraryRefreshNotifier`.
+class WatchPatchPromotionNotifier extends BaseNotifier<WatchPatchPromotion> {
+  static final WatchPatchPromotionNotifier _instance = WatchPatchPromotionNotifier._internal();
+
+  factory WatchPatchPromotionNotifier() => _instance;
+
+  WatchPatchPromotionNotifier._internal();
+
+  void promote(WatchPatchId patchId) {
+    appLogger.d('WatchPatchPromotionNotifier: promoting $patchId');
+    notify(WatchPatchPromotion(patchId));
+  }
+}

test/providers/discover_provider_test.dart

@@ -71,6 +71,8 @@ class _FakeAggregationService extends DataAggregationService {
   Set<String>? hubSucceededServerIds;
   Set<String> onDeckCancelledServerIds = const {};
   Set<String> hubCancelledServerIds = const {};
+  Set<String> onDeckFailedServerIds = const {};
+  Set<String> hubFailedServerIds = const {};
   List<MediaItem> Function() onDeckResult = () => const [];
   List<MediaHub> Function() hubsResult = () => const [];
   Future<void>? onDeckGate;
@@ -93,8 +95,10 @@ class _FakeAggregationService extends DataAggregationService {
     final items = onDeckResult();
     return (
       items: limit != null && items.length > limit ? items.sublist(0, limit) : items,
+      observedItems: [for (final item in items) (item: item, clientScope: null)],
       succeededServerIds: onDeckSucceededServerIds ?? serverIds ?? const {'server_1'},
       cancelledServerIds: onDeckCancelledServerIds,
+      failedServerIds: onDeckFailedServerIds,
     );
   }
 
@@ -112,10 +116,16 @@ class _FakeAggregationService extends DataAggregationService {
     if (started != null && !started.isCompleted) started.complete();
     final gate = hubGate;
     if (gate != null) await gate;
+    final hubs = hubsResult();
     return (
-      hubs: hubsResult(),
+      hubs: hubs,
+      observedItems: [
+        for (final hub in hubs)
+          for (final item in hub.items) (item: item, clientScope: null),
+      ],
       succeededServerIds: hubSucceededServerIds ?? serverIds ?? const {'server_1'},
       cancelledServerIds: hubCancelledServerIds,
+      failedServerIds: hubFailedServerIds,
     );
   }
 }
@@ -697,6 +707,81 @@ void main() {
     expect(aggregation.onDeckCalls, greaterThan(callsBefore));
   });
 
+  group('manual refresh reports what actually happened (#1829)', () {
+    test('a zero-success refresh reports failure while keeping the rows visible', () async {
+      aggregation.onDeckResult = () => [_item('a')];
+      aggregation.hubsResult = () => [_hub('hub-1')];
+      await provider.load();
+
+      aggregation.onDeckSucceededServerIds = const {};
+      aggregation.hubSucceededServerIds = const {};
+      aggregation.onDeckFailedServerIds = const {'server_1'};
+      aggregation.hubFailedServerIds = const {'server_1'};
+      aggregation.onDeckResult = () => const [];
+      aggregation.hubsResult = () => const [];
+
+      expect(await provider.refreshNow(), DiscoverRefreshOutcome.failed);
+      // Retained content still renders: the error is surfaced by the caller as
+      // a snackbar, never by blanking the screen.
+      expect(provider.onDeck.map((i) => i.id), ['a']);
+      expect(provider.hubs.map((h) => h.id), ['hub-1']);
+      expect(provider.errorMessage, isNull);
+    });
+
+    test('a partly failed refresh is degraded, not a success', () async {
+      aggregation.onDeckResult = () => [_item('a')];
+      aggregation.hubsResult = () => [_hub('hub-1')];
+      await provider.load();
+
+      aggregation.hubFailedServerIds = const {'server_2'};
+      expect(await provider.refreshNow(), DiscoverRefreshOutcome.degraded);
+    });
+
+    test('a cancelled refresh is not reported as a failure', () async {
+      aggregation.onDeckSucceededServerIds = const {};
+      aggregation.hubSucceededServerIds = const {};
+      aggregation.onDeckCancelledServerIds = const {'server_1'};
+      aggregation.hubCancelledServerIds = const {'server_1'};
+
+      expect(await provider.refreshNow(), DiscoverRefreshOutcome.cancelled);
+    });
+
+    test('a fully successful refresh reports success', () async {
+      aggregation.onDeckResult = () => [_item('a')];
+      aggregation.hubsResult = () => [_hub('hub-1')];
+
+      expect(await provider.refreshNow(), DiscoverRefreshOutcome.refreshed);
+    });
+
+    test('a server that failed one leg stays eligible for retry', () async {
+      aggregation.onDeckResult = () => [_item('a')];
+      aggregation.hubsResult = () => [_hub('hub-1')];
+      // Succeeded and failed are unions, so one server can appear in both;
+      // loaded ids must exclude it or syncToOnlineServers never retries.
+      aggregation.hubFailedServerIds = const {'server_1'};
+      await provider.load();
+
+      final callsBefore = aggregation.hubCalls;
+      await provider.syncToOnlineServers({'server_1'});
+      expect(aggregation.hubCalls, greaterThan(callsBefore));
+    });
+
+    test('a zero-success background refresh retains rows and stays silent', () async {
+      aggregation.onDeckResult = () => [_item('a')];
+      await provider.load();
+
+      aggregation.onDeckSucceededServerIds = const {};
+      aggregation.onDeckFailedServerIds = const {'server_1'};
+      aggregation.onDeckResult = () => const [];
+      await provider.refreshContinueWatching();
+
+      // Previously this wiped the row outright.
+      expect(provider.onDeck.map((i) => i.id), ['a']);
+      expect(provider.errorMessage, isNull);
+      expect(provider.isLoading, isFalse);
+    });
+  });
+
   test('a disrupted half is independent: on-deck commits while hubs stay loading', () async {
     aggregation.onDeckResult = () => [_item('a')];
     aggregation.hubSucceededServerIds = const {};

test/providers/watch_state_store_test.dart

@@ -5,6 +5,7 @@ import 'package:plezy/media/media_backend.dart';
 import 'package:plezy/media/media_kind.dart';
 import 'package:plezy/providers/watch_state_store.dart';
 import 'package:plezy/services/watch_state_resolver.dart';
+import 'package:plezy/utils/active_client_scope.dart';
 import 'package:plezy/utils/watch_state_notifier.dart';
 import '../test_helpers/media_items.dart';
 
@@ -22,6 +23,7 @@ WatchStateEvent _event({
   int? viewOffset,
   List<String> parentChain = const [],
   String mediaType = 'movie',
+  bool serverAcknowledged = false,
 }) {
   return WatchStateEvent(
     itemId: itemId,
@@ -32,9 +34,13 @@ WatchStateEvent _event({
     mediaType: mediaType,
     isNowWatched: isNowWatched,
     viewOffset: viewOffset,
+    serverAcknowledged: serverAcknowledged,
+    patchId: WatchPatchId.session(++_testSequence),
   );
 }
 
+var _testSequence = 0;
+
 final _episode = testMediaItem(
   id: 'episode-1',
   backend: MediaBackend.jellyfin,
@@ -231,4 +237,447 @@ void main() {
     store.setActiveClientScopesByServer({'jf-machine': 'jf-machine/user-b'});
     expect(store.apply(_episode).isWatched, isFalse);
   });
+
+  group('authoritative observations supersede acknowledged patches (#1829)', () {
+    /// The reporter's sequence: the Mac pauses an episode, another device
+    /// advances the server, and a refresh must render the server's position
+    /// rather than the Mac's.
+    test('a fetch that started after the patch renders the server offset', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+      expect(store.apply(fresh).viewOffsetMs, 1800000, reason: 'precondition: the stale patch still wins');
+
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.apply(fresh).viewOffsetMs, 2700000);
+      expect(store.patchForItem(fresh), isNull);
+    });
+
+    test('a patch recorded during the fetch survives its observation', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      // Watermark captured before the request, patch emitted while it is in
+      // flight: the response is older than the local action and must lose.
+      final watermark = store.observationWatermark;
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      store.recordObservations([(item: fresh, clientScope: null)], watermark: watermark, epoch: store.observationEpoch);
+
+      expect(store.apply(fresh).viewOffsetMs, 1800000);
+    });
+
+    test('an unacknowledged patch is never superseded by a read', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      // An offline write still owed to the server must outlive any fetch.
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.apply(fresh).viewOffsetMs, 1800000);
+    });
+
+    test('an observation from another store epoch is rejected', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      // A request issued before a profile switch, landing after it: its
+      // watermark belongs to a store that no longer exists.
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: Object(),
+      );
+
+      expect(store.apply(fresh).viewOffsetMs, 1800000);
+    });
+
+    test('an older response cannot lower an observation already recorded', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      final stale = store.observationWatermark;
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+      // A concurrent request that started earlier completes last.
+      store.recordObservations([(item: fresh, clientScope: null)], watermark: stale, epoch: store.observationEpoch);
+
+      expect(store.apply(fresh).viewOffsetMs, 2700000, reason: 'max() per key, so response order cannot matter');
+    });
+
+    test('suppressing an item does not expose an older ancestor patch', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      // Season marked unwatched first, then the episode played: without the
+      // ancestor barrier, suppressing the episode entry would let the older
+      // season patch win and render unwatched/0 — worse than either value.
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.unwatched,
+          isNowWatched: false,
+          itemId: 'season-1',
+          mediaType: 'season',
+          serverAcknowledged: true,
+        ),
+      );
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000, viewCount: 0);
+
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.patchForItem(fresh), isNull);
+      expect(store.apply(fresh).viewOffsetMs, 2700000);
+    });
+
+    test('an ancestor mark newer than the observation still reaches descendants', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+      // The container action happens after the fetch, so it must survive.
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.watched,
+          isNowWatched: true,
+          itemId: 'season-1',
+          mediaType: 'season',
+          serverAcknowledged: true,
+        ),
+      );
+
+      expect(store.apply(fresh).isWatched, isTrue);
+    });
+
+    test('an unobserved sibling still sees the container mark', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.watched,
+          isNowWatched: true,
+          itemId: 'season-1',
+          mediaType: 'season',
+          serverAcknowledged: true,
+        ),
+      );
+      final observed = _episode.copyWith(viewCount: 0);
+      final sibling = testMediaItem(
+        id: 'episode-2',
+        backend: MediaBackend.jellyfin,
+        kind: MediaKind.episode,
+        parentId: 'season-1',
+        grandparentId: 'show-1',
+        serverId: 'jf-machine',
+      );
+
+      store.recordObservations(
+        [(item: observed, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.apply(observed).isWatched, isFalse, reason: 'the observed child yields to the server');
+      expect(store.apply(sibling).isWatched, isTrue, reason: 'the unobserved sibling keeps the container mark');
+    });
+
+    test('an observation under one client scope cannot suppress another scope', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+      store.setActiveClientScopesByServer({'jf-machine': 'jf-machine/user-b'});
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          cacheServerId: 'jf-machine/user-b',
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      // User A's request returning after the switch back to B: the public
+      // server id is shared, so keying by it would suppress B's watch state.
+      store.recordObservations(
+        [(item: fresh, clientScope: 'jf-machine/user-a')],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.apply(fresh).viewOffsetMs, 1800000);
+    });
+
+    test('a profile switch clears observations along with patches', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+      store.setActiveProfileId('profile-2');
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          itemId: 'episode-1',
+          viewOffset: 900000,
+          serverAcknowledged: true,
+        ),
+      );
+
+      expect(store.apply(fresh).viewOffsetMs, 900000, reason: 'a stale observation must not suppress the new patch');
+    });
+  });
+
+  group('promotion settles an owed write (#1829)', () {
+    test('a promoted patch becomes suppressible', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      final patchId = WatchStateNotifier().notifyProgress(item: _episode, viewOffset: 1800000, duration: 3600000);
+      await Future<void>.delayed(Duration.zero);
+      final fresh = _episode.copyWith(viewOffsetMs: 2700000, durationMs: 3600000);
+
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+      expect(store.apply(fresh).viewOffsetMs, 1800000, reason: 'unacknowledged, so the read must not retire it');
+
+      WatchPatchPromotionNotifier().promote(patchId!);
+      await Future<void>.delayed(Duration.zero);
+
+      // Promotion assigns a fresh sequence, so the observation that preceded
+      // the write cannot suppress it; a later one can.
+      expect(store.apply(fresh).viewOffsetMs, 1800000);
+      store.recordObservations(
+        [(item: fresh, clientScope: null)],
+        watermark: store.observationWatermark,
+        epoch: store.observationEpoch,
+      );
+      expect(store.apply(fresh).viewOffsetMs, 2700000);
+    });
+
+    test('promoting an unknown id is a no-op', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+
+      await _emit(_event(changeType: WatchStateChangeType.progressUpdate, isNowWatched: false, viewOffset: 1800000));
+      WatchPatchPromotionNotifier().promote(WatchPatchId.session(999999));
+      await Future<void>.delayed(Duration.zero);
+
+      expect(store.patchForGlobalKey('jf-machine:item-1')?.viewOffsetMs, 1800000);
+    });
+  });
+
+  // The reporter's setup: Plex on macOS, where the client's cache scope is a
+  // profile scope that differs from the public server id. The observation and
+  // the patch must land on the same key or suppression silently never fires.
+  group("the reporter's Plex handoff (#1829)", () {
+    final plexScope = buildPlexProfileScopeId(serverId: ServerId('plex-machine'), profileId: 'profile-a');
+    final plexEpisode = testMediaItem(
+      id: 'episode-1',
+      backend: MediaBackend.plex,
+      kind: MediaKind.episode,
+      serverId: 'plex-machine',
+      durationMs: 3600000,
+    );
+
+    test('a refreshed row overrides the offset the Mac paused at', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+      // Exactly how ProfileSessionScreen wires the map: serverId -> the
+      // client's own cacheServerId.
+      store.setActiveClientScopesByServer({'plex-machine': plexScope});
+
+      // The Mac pauses with 30 minutes left. Plex persists position on every
+      // timeline report, so the tracker acknowledges it.
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          serverId: 'plex-machine',
+          itemId: 'episode-1',
+          cacheServerId: plexScope,
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+
+      // Precondition: without an observation the local patch pins the row.
+      // This is the bug as reported.
+      final serverRow = plexEpisode.copyWith(viewOffsetMs: 2700000);
+      expect(store.apply(serverRow).viewOffsetMs, 1800000);
+
+      // The iPad advances the server to 15 minutes left; Refresh reads it back.
+      final watermark = store.observationWatermark;
+      store.recordObservations(
+        [(item: serverRow, clientScope: plexScope)],
+        watermark: watermark,
+        epoch: store.observationEpoch,
+      );
+
+      expect(store.apply(serverRow).viewOffsetMs, 2700000);
+    });
+
+    test('a pre-refresh pause still pins, so a later local action is not lost', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+      store.setActiveClientScopesByServer({'plex-machine': plexScope});
+
+      final serverRow = plexEpisode.copyWith(viewOffsetMs: 2700000);
+      final watermark = store.observationWatermark;
+      store.recordObservations(
+        [(item: serverRow, clientScope: plexScope)],
+        watermark: watermark,
+        epoch: store.observationEpoch,
+      );
+
+      // Paused after that read: newer than the watermark, so it must survive.
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          serverId: 'plex-machine',
+          itemId: 'episode-1',
+          cacheServerId: plexScope,
+          viewOffset: 3000000,
+          serverAcknowledged: true,
+        ),
+      );
+
+      expect(store.apply(serverRow).viewOffsetMs, 3000000);
+    });
+
+    test('an unscoped store neither pins nor mis-suppresses', () async {
+      final store = WatchStateStore();
+      addTearDown(store.dispose);
+      // DownloadMetadataStore replaces the scope map with only the servers it
+      // has downloads for, so a Plex server without downloads can end up
+      // unscoped. The patch keys off the event's own cacheServerId, so both
+      // sides miss together rather than leaving a stale pin behind.
+      store.setActiveClientScopesByServer(const {});
+
+      await _emit(
+        _event(
+          changeType: WatchStateChangeType.progressUpdate,
+          isNowWatched: false,
+          serverId: 'plex-machine',
+          itemId: 'episode-1',
+          cacheServerId: plexScope,
+          viewOffset: 1800000,
+          serverAcknowledged: true,
+        ),
+      );
+
+      final serverRow = plexEpisode.copyWith(viewOffsetMs: 2700000);
+      expect(store.apply(serverRow).viewOffsetMs, 2700000);
+    });
+  });
 }

test/screens/discover_screen_test.dart

@@ -651,8 +651,11 @@ class _FakeMediaServerClient implements MediaServerClient {
   Future<List<MediaItem>> fetchContinueWatching({int? count = 20}) async => continueWatching;
 
   @override
-  Future<List<MediaHub>> fetchGlobalHubs({int limit = defaultHubPreviewLimit, bool includePlaybackHubs = true}) async =>
+  Future<List<MediaHub>> fetchGlobalHubs({
-      hubs;
+    int limit = defaultHubPreviewLimit,
+    bool includePlaybackHubs = true,
+    HubFetchDiagnostics? diagnostics,
+  }) async => hubs;
 
   @override
   dynamic noSuchMethod(Invocation invocation) => super.noSuchMethod(invocation);
@@ -692,7 +695,11 @@ class _GatedHubsFakeClient implements MediaServerClient {
   Future<List<MediaLibrary>> fetchLibraries() async => const [];
 
   @override
-  Future<List<MediaHub>> fetchGlobalHubs({int limit = defaultHubPreviewLimit, bool includePlaybackHubs = true}) {
+  Future<List<MediaHub>> fetchGlobalHubs({
+    int limit = defaultHubPreviewLimit,
+    bool includePlaybackHubs = true,
+    HubFetchDiagnostics? diagnostics,
+  }) {
     hubCalls++;
     final gate = Completer<List<MediaHub>>();
     _gates.add(gate);

test/services/data_aggregation_bridge_test.dart

@@ -126,6 +126,7 @@ class _GatedHubsClient implements MediaServerClient {
     int limit = defaultHubPreviewLimit,
     bool includePlaybackHubs = true,
     MediaKind? libraryKind,
+    HubFetchDiagnostics? diagnostics,
   }) {
     started.add(libraryId);
     return (_gates[libraryId] = Completer<List<MediaHub>>()).future;

test/services/offline_watch_sync_service_test.dart

@@ -981,7 +981,7 @@ void main() {
       mgr.debugRegisterClientForTesting(clientA);
 
       final queuedScope = await svc.queueMarkWatched(serverId: ServerId('plex-machine'), itemId: 'item-1');
-      expect(queuedScope, clientA.profileScopeId);
+      expect(queuedScope.clientScopeId, clientA.profileScopeId);
       expect((await db.getPendingWatchActions()).single.clientScopeId, clientA.profileScopeId);
 
       await svc.syncPendingItems();
@@ -1044,7 +1044,7 @@ void main() {
       final returnedScope = await svc.queueMarkWatched(serverId: ServerId('jf-machine'), itemId: 'item-1');
 
       final queued = await db.getPendingWatchActions();
-      expect(returnedScope, 'jf-machine/user-a');
+      expect(returnedScope.clientScopeId, 'jf-machine/user-a');
       expect(queued.single.clientScopeId, 'jf-machine/user-a');
     });
 
@@ -1145,7 +1145,7 @@ void main() {
       final returnedScope = await svc.queueMarkWatched(serverId: ServerId('jf-machine'), itemId: 'item-1');
 
       final queued = await db.getPendingWatchActions();
-      expect(returnedScope, 'jf-machine/user-b');
+      expect(returnedScope.clientScopeId, 'jf-machine/user-b');
       expect(queued.single.clientScopeId, 'jf-machine/user-b');
     });