feat(voice): all microphone buttons use the server's speech recognition

- New useVoiceInput hook (lib/voice.ts): "server" records with MediaRecorder
  and sends to /voice/transcribe, "browser" keeps the live recognition,
  "off" or missing browser support hides the microphone.
- Quick dump bar, mobile capture sheet and hold-to-record use it; live
  waves from the real input level (VoiceWave) and "Wird erkannt…".
- Recordings stop after STT_MAX_SECONDS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

frontend/src/lib/__tests__/useVoiceInput.test.tsx

@@ -0,0 +1,133 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { act, renderHook, waitFor } from '@testing-library/react';
+
+const getAuthStatus = vi.fn();
+const transcribe = vi.fn();
+vi.mock('../api', () => ({ api: { getAuthStatus: (...a: any[]) => getAuthStatus(...a), transcribe: (...a: any[]) => transcribe(...a) } }));
+
+import { useVoiceInput } from '../useVoiceInput';
+import { resetVoiceConfig } from '../voice';
+
+class FakeMediaRecorder {
+  static isTypeSupported = (t: string) => t === 'audio/webm;codecs=opus';
+  state: 'inactive' | 'recording' = 'inactive';
+  mimeType = 'audio/webm;codecs=opus';
+  ondataavailable: ((e: { data: Blob }) => void) | null = null;
+  onstop: (() => void) | null = null;
+  constructor(public stream: any) {}
+  start() {
+    this.state = 'recording';
+  }
+  stop() {
+    this.state = 'inactive';
+    this.ondataavailable?.({ data: new Blob(['audio'], { type: this.mimeType }) });
+    this.onstop?.();
+  }
+}
+
+const stopTrack = vi.fn();
+
+beforeEach(() => {
+  resetVoiceConfig();
+  getAuthStatus.mockReset();
+  transcribe.mockReset();
+  stopTrack.mockReset();
+  (window as any).MediaRecorder = FakeMediaRecorder;
+  Object.defineProperty(navigator, 'mediaDevices', {
+    configurable: true,
+    value: { getUserMedia: vi.fn(async () => ({ getTracks: () => [{ stop: stopTrack }] })) },
+  });
+});
+
+afterEach(() => {
+  delete (window as any).MediaRecorder;
+  delete (window as any).SpeechRecognition;
+  vi.useRealTimers();
+});
+
+describe('useVoiceInput', () => {
+  it('server mode: records, sends the audio to the own server and returns its text', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    transcribe.mockResolvedValue('Milch kaufen');
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    expect(result.current.available).toBe(true);
+
+    await act(async () => {
+      await result.current.start();
+    });
+    expect(result.current.state).toBe('listening');
+
+    await act(async () => {
+      await result.current.stop();
+    });
+    expect(transcribe).toHaveBeenCalledTimes(1);
+    expect((transcribe.mock.calls[0][0] as Blob).type).toBe('audio/webm;codecs=opus');
+    expect(onText).toHaveBeenCalledWith('Milch kaufen', true);
+    expect(result.current.state).toBe('idle');
+    expect(stopTrack).toHaveBeenCalled(); // microphone released
+  });
+
+  it('server mode: cancel throws the recording away', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    await act(async () => {
+      await result.current.start();
+    });
+    act(() => result.current.cancel());
+    expect(transcribe).not.toHaveBeenCalled();
+    expect(onText).not.toHaveBeenCalled();
+    expect(result.current.state).toBe('idle');
+  });
+
+  it('server mode: a failing server reports an error instead of text', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    transcribe.mockRejectedValue(new Error('Spracherkennung fehlgeschlagen'));
+    const onText = vi.fn();
+    const onError = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText, onError }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    await act(async () => {
+      await result.current.start();
+    });
+    await act(async () => {
+      await result.current.stop();
+    });
+    expect(onError).toHaveBeenCalledWith('Spracherkennung fehlgeschlagen');
+    expect(onText).not.toHaveBeenCalled();
+    expect(result.current.state).toBe('idle');
+  });
+
+  it('server mode: stops by itself after the maximum length', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 5 } });
+    transcribe.mockResolvedValue('Zu lang');
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    vi.useFakeTimers();
+    await act(async () => {
+      await result.current.start();
+    });
+    await act(async () => {
+      vi.advanceTimersByTime(5000);
+    });
+    vi.useRealTimers();
+    await waitFor(() => expect(onText).toHaveBeenCalledWith('Zu lang', true));
+  });
+
+  it('hides the microphone when switched off, and without browser recognition', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'off', maxSeconds: 600 } });
+    const off = renderHook(() => useVoiceInput({ onText: vi.fn() }));
+    await waitFor(() => expect(off.result.current.mode).toBe('off'));
+    expect(off.result.current.available).toBe(false);
+
+    resetVoiceConfig();
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'browser', maxSeconds: 600 } });
+    const none = renderHook(() => useVoiceInput({ onText: vi.fn() }));
+    await waitFor(() => expect(none.result.current.mode).toBe('unsupported'));
+    expect(none.result.current.available).toBe(false);
+  });
+});