fix(settings): valid HTML in setting tiles; start uploads in an effect (#291)

SettingTile rendered its description in a <p>, and many tiles pass
blocks (badges, Box, Text), so Settings had <div>/<p> inside <p> and
React warned on every page. Render it as a <div>. UploadCardRenderer
called startUpload() during render, which now updates the desktop
taskbar-progress hook mid-render; start it from an effect instead.

Refs #291

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5

src/app/components/CallEmbedProvider.tsx

@@ -86,7 +86,6 @@ import { CallAvatarAnimation } from '../styles/Animations.css';
 import { webRTCSupported } from '../utils/rtc';
 import { CallsUnavailableMessage } from '../features/call/CallsUnavailable';
 import { zIndices } from '../styles/zIndex';
-import { useDenoiseNotices } from '../hooks/useDenoiseNotices';
 
 const PIP_MIN_W = 200;
 const PIP_MIN_H = 112;
@@ -794,7 +793,6 @@ function CallUtils({ embed, joined }: { embed: CallEmbed; joined: boolean }) {
   useCallPolicyRevokedToast(embed, joined);
   useCallEndedToast(embed);
   useScreenshareNotices(embed);
-  useDenoiseNotices(embed);
   // [Gitea #43] The in-frame screenshare button follows the same room policy
   // as the host bar's (hidden where the server would refuse the share).
   const { allowScreenshare } = useRoomCallPolicy(embed.room);

src/app/components/setting-tile/SettingTile.tsx

@@ -43,7 +43,9 @@ export function SettingTile({ title, description, before, after, children }: Set
           </Text>
         )}
         {description && (
-          <Text className={BreakWord} size="T200" priority="300">
+          // [Gitea #291] A <div>, not a <p>: many tiles pass blocks (badges,
+          // <Box>, <Text>) as the description, which is invalid inside a <p>.
+          <Text as="div" className={BreakWord} size="T200" priority="300">
             {description}
           </Text>
         )}

src/app/components/upload-card/UploadCardRenderer.tsx

@@ -310,9 +310,13 @@ export function UploadCardRenderer({
   const { file } = upload;
   const fileSizeExceeded = file.size >= allowSize;
 
-  if (upload.status === UploadStatus.Idle && !fileSizeExceeded) {
-    startUpload();
-  }
+  // [Gitea #291] Start in an effect, not during render: the desktop taskbar
+  // progress hook subscribes to the upload atoms, so a render-time start
+  // updated another component mid-render (React warning on every image send).
+  const idle = upload.status === UploadStatus.Idle;
+  useEffect(() => {
+    if (idle && !fileSizeExceeded) startUpload();
+  }, [idle, fileSizeExceeded, startUpload]);
 
   const handleSpoiler = (marked: boolean) => {
     setMetadata(fileItem, { ...metadata, markedAsSpoiler: marked });

src/app/hooks/useDenoiseNotices.ts

@@ -1,38 +0,0 @@
-import { useEffect, useRef } from 'react';
-import { useSetAtom } from 'jotai';
-import { Icons } from 'folds';
-import { CallEmbed } from '../plugins/call';
-import { toastQueueAtom } from '../state/toast';
-import { useSetting } from '../state/hooks/settings';
-import { settingsAtom } from '../state/settings';
-import { denoiseNotice, DenoiseStateReport } from '../utils/denoiseNotice';
-
-/**
- * [Gitea #260] Tell the user when ML noise suppression didn't attach (raw mic)
- * or fell back to another model. The fork's `io.lotus.denoise_state` used to
- * go unanswered ("unsupported action"), so failures were silent.
- */
-export function useDenoiseNotices(embed: CallEmbed): void {
-  const setToast = useSetAtom(toastQueueAtom);
-  const [mode] = useSetting(settingsAtom, 'callNoiseSuppression');
-  const [model] = useSetting(settingsAtom, 'callDenoiseModel');
-  const latest = useRef({ mode, model });
-  latest.current = { mode, model };
-  useEffect(
-    () =>
-      embed.listenAction<{ data?: DenoiseStateReport }>('io.lotus.denoise_state', (evt) => {
-        if (latest.current.mode !== 'ml') return;
-        const body = denoiseNotice(latest.current.model, evt.detail?.data);
-        if (!body) return;
-        setToast({
-          id: `denoise-state-${Date.now()}`,
-          iconSrc: Icons.Mic,
-          displayName: 'Noise suppression',
-          body,
-          roomName: embed.room.name ?? 'Voice call',
-          roomId: embed.roomId,
-        });
-      }),
-    [embed, setToast],
-  );
-}

src/app/pages/client/ClientNonUIFeatures.tsx

@@ -1060,6 +1060,32 @@ function TauriUpdateFeature() {
   return null;
 }
 
+function LotusDenoiseFeature() {
+  const setToast = useSetAtom(toastQueueAtom);
+
+  useEffect(() => {
+    const handleMessage = (event: MessageEvent) => {
+      if (event.data?.type === 'lotus-denoise-status') {
+        const { active, error } = event.data;
+        if (!active) {
+          setToast({
+            id: `denoise-fail-${Date.now()}`,
+            displayName: 'Audio Quality',
+            body: `ML Noise Suppression failed: ${error || 'Unknown error'}. Falling back to raw mic.`,
+            roomName: 'System',
+            roomId: '',
+          });
+        }
+      }
+    };
+
+    window.addEventListener('message', handleMessage);
+    return () => window.removeEventListener('message', handleMessage);
+  }, [setToast]);
+
+  return null;
+}
+
 // Registers the global `?` shortcut (ignored while typing) and renders the
 // keyboard-shortcuts help dialog. Headless — the dialog self-gates on its atom.
 function KeyboardShortcutsFeature() {
@@ -1141,6 +1167,7 @@ export function ClientNonUIFeatures({ children }: ClientNonUIFeaturesProps) {
       <CloseBehaviorPrompt />
       <ConsentRequiredPrompt />
       <OutboxFeature />
+      <LotusDenoiseFeature />
       <DeepLinkNavigator />
       <KeyboardShortcutsFeature />
       {children}

src/app/utils/denoiseNotice.test.ts

@@ -1,17 +0,0 @@
-import { test } from 'node:test';
-import assert from 'node:assert/strict';
-import { denoiseNotice } from './denoiseNotice';
-
-test('denoiseNotice: says when ML is off or fell back (#260)', () => {
-  assert.equal(denoiseNotice('deepfilternet', { active: true, model: 'deepfilternet' }), undefined);
-  assert.match(
-    denoiseNotice('deepfilternet', { active: true, model: 'rnnoise' }) ?? '',
-    /^DeepFilterNet 3 couldn't start, so noise suppression is using RNNoise instead\.$/,
-  );
-  assert.match(
-    denoiseNotice('rnnoise', { active: false, error: 'boom' }) ?? '',
-    /unfiltered audio/,
-  );
-  assert.equal(denoiseNotice('rnnoise', undefined), undefined);
-  assert.equal(denoiseNotice('rnnoise', { model: 'x' }), undefined);
-});

src/app/utils/denoiseNotice.ts

@@ -1,29 +0,0 @@
-/**
- * [Gitea #260] The fork reports what noise suppression actually runs as
- * `io.lotus.denoise_state { active, model, error? }`. Pure: the toast text for
- * a report, or undefined when what runs is what the user picked.
- */
-export type DenoiseStateReport = { active?: unknown; model?: unknown; error?: unknown };
-
-const MODEL_NAMES: Record<string, string> = {
-  rnnoise: 'RNNoise',
-  speex: 'Speex',
-  dtln: 'DTLN',
-  deepfilternet: 'DeepFilterNet 3',
-};
-const modelName = (m: unknown): string =>
-  typeof m === 'string' ? (MODEL_NAMES[m] ?? m) : 'the selected model';
-
-export function denoiseNotice(
-  requestedModel: string,
-  report: DenoiseStateReport | undefined,
-): string | undefined {
-  if (!report || typeof report.active !== 'boolean') return undefined;
-  if (!report.active) {
-    return `ML noise suppression couldn't start, so your microphone is sending unfiltered audio. You can pick another model in Settings → Calls.`;
-  }
-  if (typeof report.model === 'string' && report.model !== requestedModel) {
-    return `${modelName(requestedModel)} couldn't start, so noise suppression is using ${modelName(report.model)} instead.`;
-  }
-  return undefined;
-}