feat(briefing): stall watchdog + pull-to-refresh; chat attaches to in-flight streams

Briefing chat now uses the same StreamIterator + 45s per-event timeout
as the main chat provider, so a dropped SSE socket no longer leaves
isBriefingStreaming stuck true. Adds refreshMessages() that unfreezes
state when the server-side message is complete, wired to a new
pull-to-refresh gesture and the app lifecycle-resume hook.

Chat provider gains attachToGeneration() — safe to call unconditionally
on screen init, so landing on a conversation that's already mid-stream
(e.g. the /news discuss button, which now creates a conv and auto-kicks
generation server-side) picks up live tokens instead of freezing on the
placeholder.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

lib/providers/chat_provider.dart

@@ -118,6 +118,120 @@ class MessagesNotifier extends AsyncNotifier<List<Message>> {
     }
   }
 
+  /// Attach to an already-running generation for this conversation.
+  ///
+  /// Used when the chat screen lands on a conversation that was started by
+  /// something other than a direct user message — e.g. the /news discuss
+  /// button, which creates a conversation on the backend and auto-kicks a
+  /// generation before navigating. Without this the stream runs to
+  /// completion invisibly and the screen only shows the final persisted
+  /// message after a manual refresh.
+  ///
+  /// Safe to call unconditionally on screen init: no-ops when there is no
+  /// generating assistant message. Mirrors the web chat store's
+  /// reconnectIfGenerating() helper.
+  Future<void> attachToGeneration() async {
+    final convId = _convId;
+    final repo = ref.read(chatRepositoryProvider);
+
+    // Make sure we're looking at fresh server state before deciding whether
+    // to attach. The provider's build() fetches once; if the conversation
+    // was seeded via a POST that happened between build() and this call the
+    // generating placeholder won't be in our in-memory list yet.
+    try {
+      final (_, fresh) = await repo.getMessages(convId);
+      state = AsyncData(fresh);
+    } catch (_) {
+      // If we can't load messages we can't attach either — bail cleanly.
+      return;
+    }
+
+    if (ref.read(isStreamingProvider(convId))) return;
+    final msgs = state.value ?? const <Message>[];
+    final hasGeneratingAssistant = msgs.any(
+      (m) => m.role == MessageRole.assistant && m.status == 'generating',
+    );
+    if (!hasGeneratingAssistant) return;
+
+    ref.read(isStreamingProvider(convId).notifier).state = true;
+
+    const stallTimeout = Duration(seconds: 45);
+    bool streamedContent = false;
+    final iter = StreamIterator(repo.streamGeneration(convId));
+    try {
+      while (await iter.moveNext().timeout(stallTimeout)) {
+        final event = iter.current;
+        if (event is ChatTextChunk) {
+          streamedContent = true;
+          ref.read(streamingStatusProvider(convId).notifier).state = '';
+          final cur = state.requireValue;
+          if (cur.isEmpty) continue;
+          // Route text chunks into the generating assistant message. The
+          // last message is usually the placeholder, but tool-call fan-in
+          // means we can't rely on that universally.
+          final idx = _findGeneratingAssistantIndex(cur);
+          if (idx < 0) continue;
+          final updated = cur[idx].copyWith(content: cur[idx].content + event.text);
+          state = AsyncData([
+            ...cur.sublist(0, idx),
+            updated,
+            ...cur.sublist(idx + 1),
+          ]);
+        } else if (event is ChatStatusUpdate) {
+          ref.read(streamingStatusProvider(convId).notifier).state = event.status;
+        } else if (event is ChatToolCall) {
+          final cur = state.requireValue;
+          if (cur.isEmpty) continue;
+          final idx = _findGeneratingAssistantIndex(cur);
+          if (idx < 0) continue;
+          final nextCalls = [...?cur[idx].toolCalls, event.toolCall];
+          final updated = cur[idx].copyWith(toolCalls: nextCalls);
+          state = AsyncData([
+            ...cur.sublist(0, idx),
+            updated,
+            ...cur.sublist(idx + 1),
+          ]);
+        }
+      }
+    } on TimeoutException {
+      // Stall — fall through to polling.
+    } catch (_) {
+      // Stream failed — fall through to polling.
+    } finally {
+      await iter.cancel();
+    }
+
+    try {
+      for (var attempt = 0; attempt < 20; attempt++) {
+        if (attempt > 0) await Future.delayed(const Duration(seconds: 2));
+        final (_, fresh) = await repo.getMessages(convId);
+        final done = fresh.any(
+          (m) => m.role == MessageRole.assistant && m.status != 'generating',
+        );
+        final polledHasContent = fresh.any(
+          (m) => m.role == MessageRole.assistant && m.content.isNotEmpty,
+        );
+        if (!streamedContent || done || polledHasContent) {
+          state = AsyncData(fresh);
+        }
+        if (done) break;
+      }
+    } catch (_) {
+      // Give up silently — user can pull-to-refresh.
+    } finally {
+      ref.read(isStreamingProvider(convId).notifier).state = false;
+      ref.read(streamingStatusProvider(convId).notifier).state = '';
+    }
+  }
+
+  int _findGeneratingAssistantIndex(List<Message> msgs) {
+    for (var i = msgs.length - 1; i >= 0; i--) {
+      final m = msgs[i];
+      if (m.role == MessageRole.assistant && m.status == 'generating') return i;
+    }
+    return -1;
+  }
+
   Future<void> sendMessage(String content) async {
     final convId = _convId;
     final repo = ref.read(chatRepositoryProvider);