feat(desktop): LLM 기본 모델 qwen3:4b → gemma4:e4b 전면 전환 + think:false 안전장치

qwen3:4b가 reasoning 모델이라 <think>...</think> 블록을 길게 생성 →
stripReasoningBlocks 후 빈 문자열 → 원본 transcript fallback으로 끝나면서
LLM refine이 42초 걸리는 병목 발견. Google Gemma 4 e4b(4.5B effective
params, 2026-04-02 릴리스)로 교체. non-reasoning 기본 + Ollama v0.20+
think: false 파라미터로 2중 방어.

실측 결과: 받아쓰기 한 사이클 51.4s → 5.5s (9.3배 빠름).
  STT 500ms + LLM 3,925ms + insert 1,092ms.
refine 품질 정상 동작 확인: "테스트하는 중입니다" → "테스트하고 있습니다".

- LocalLLMService: 3개 fallback 기본값 변경(generate / streamGenerate /
  chatStream) + Ollama 요청 body에 think: false 명시 추가. non-reasoning
  모델은 무시, reasoning 모델은 thinking 토큰 차단. NO_THINK 주석을
  legacy 설명으로 업데이트 — qwen3/deepseek-r1 수동 선택자를 위한 3중
  방어(/no_think + think:false + stripReasoningBlocks) 명시.
- OnboardingModal / OllamaGuideModal: pull 명령어 갱신
- 테스트 fixture 갱신
- 12개 i18n locale JSON: settings.ollamaHint / ollama.step2.alt 키 업데이트
  (qwen3:4b → gemma4:e4b, qwen3:8b → gemma4:26b)
- 10개 site i18n locale TS + HowItWorks.tsx 파이프라인 시각화 — detail
  문자열 'qwen3 / llama3 / gemma3' → 'gemma4 / llama3.2 / phi4',
  파이프라인 라벨 'qwen3:4b @ localhost' → 'gemma4:e4b @ localhost'
- 설계서 00 LLMConfig 기본값 + CONFIG_DEFAULTS
- 설계서 05: 6개 API 스키마 예시, 2개 OllamaClient 코드 예시, LLM 모델
  추천 표 재정렬(gemma4:e4b 최상위, qwen3는 reasoning 경고와 함께 후순위),
  권장 JSON 설정에 think:false 추가
- phase-14 meeting mode 컨텍스트 윈도우 표 갱신
- V2-5 Mac 부트스트랩 가이드 pull 커맨드 갱신
- project_status.md Part 7 전체 섹션 추가

site/src/i18n/locales/de.ts

@@ -47,7 +47,7 @@ export const de: Translations = {
     steps: [
       { label: 'EINGABE', title: 'Hotkey drucken', description: 'Right Alt (oder Ihre benutzerdefinierte Taste) gedruckt halten zum Aufnehmen. Doppeltippen fur KI-Modus. Umschalten fur Freisprechen.', detail: 'Halten / Umschalten / Doppeldruck' },
       { label: 'TRANSKRIBIEREN', title: 'Whisper STT', description: 'Lokale faster-whisper Engine konvertiert 16kHz PCM-Audio in Echtzeit zu Text. GPU-Beschleunigung unterstutzt.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'VERARBEITEN', title: 'Ollama LLM', description: 'Lokales LLM korrigiert Grammatik, passt den Ton an, ubersetzt oder fasst zusammen. Benutzerdefinierte Anweisungen unterstutzt.', detail: 'qwen3 / llama3 / gemma3 / benutzerdefiniert' },
+      { label: 'VERARBEITEN', title: 'Ollama LLM', description: 'Lokales LLM korrigiert Grammatik, passt den Ton an, ubersetzt oder fasst zusammen. Benutzerdefinierte Anweisungen unterstutzt.', detail: 'gemma4 / llama3.2 / phi4 / benutzerdefiniert' },
       { label: 'AUSGABE', title: 'Auto-Einfugen', description: 'Korrigierter Text wird an der Cursorposition in jeder App eingefugt. Notepad, VS Code, Chrome, Slack, uberall.', detail: 'Zwischenablage + Ctrl+V / 100ms Latenz' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/en.ts

@@ -47,7 +47,7 @@ export const en: Translations = {
     steps: [
       { label: 'INPUT', title: 'Press Hotkey', description: 'Hold Right Alt (or your custom key) to start recording. Double-tap for AI mode. Toggle for hands-free.', detail: 'Hold-to-talk / Toggle / Double-press' },
       { label: 'TRANSCRIBE', title: 'Whisper STT', description: 'Local faster-whisper engine converts 16kHz PCM audio to text in real-time. GPU acceleration supported.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'PROCESS', title: 'Ollama LLM', description: 'Local LLM polishes grammar, adjusts tone, translates, or summarizes. Custom instructions supported.', detail: 'qwen3 / llama3 / gemma3 / custom' },
+      { label: 'PROCESS', title: 'Ollama LLM', description: 'Local LLM polishes grammar, adjusts tone, translates, or summarizes. Custom instructions supported.', detail: 'gemma4 / llama3.2 / phi4 / custom' },
       { label: 'OUTPUT', title: 'Auto Insert', description: 'Polished text is pasted at your cursor position in any app. Notepad, VS Code, Chrome, Slack, anywhere.', detail: 'Clipboard + Ctrl+V / 100ms latency' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/es.ts

@@ -47,7 +47,7 @@ export const es: Translations = {
     steps: [
       { label: 'ENTRADA', title: 'Pulsa la tecla', description: 'Manten Right Alt (o tu tecla personalizada) para grabar. Doble toque para modo IA. Alternar para manos libres.', detail: 'Mantener / Alternar / Doble pulsacion' },
       { label: 'TRANSCRIBIR', title: 'Whisper STT', description: 'El motor local faster-whisper convierte audio PCM 16kHz a texto en tiempo real. Compatible con aceleracion GPU.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'PROCESAR', title: 'Ollama LLM', description: 'El LLM local pule la gramatica, ajusta el tono, traduce o resume. Instrucciones personalizadas soportadas.', detail: 'qwen3 / llama3 / gemma3 / personalizado' },
+      { label: 'PROCESAR', title: 'Ollama LLM', description: 'El LLM local pule la gramatica, ajusta el tono, traduce o resume. Instrucciones personalizadas soportadas.', detail: 'gemma4 / llama3.2 / phi4 / personalizado' },
       { label: 'SALIDA', title: 'Insercion automatica', description: 'El texto pulido se pega en la posicion del cursor en cualquier app. Bloc de notas, VS Code, Chrome, Slack, donde sea.', detail: 'Portapapeles + Ctrl+V / 100ms latencia' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/fr.ts

@@ -47,7 +47,7 @@ export const fr: Translations = {
     steps: [
       { label: 'ENTREE', title: 'Appuyez sur la touche', description: 'Maintenez Right Alt (ou votre touche personnalisee) pour enregistrer. Double appui pour le mode IA. Basculez pour mains libres.', detail: 'Maintenir / Basculer / Double appui' },
       { label: 'TRANSCRIRE', title: 'Whisper STT', description: 'Le moteur local faster-whisper convertit l\'audio PCM 16kHz en texte en temps reel. Acceleration GPU supportee.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'TRAITER', title: 'Ollama LLM', description: 'Le LLM local corrige la grammaire, ajuste le ton, traduit ou resume. Instructions personnalisees supportees.', detail: 'qwen3 / llama3 / gemma3 / personnalise' },
+      { label: 'TRAITER', title: 'Ollama LLM', description: 'Le LLM local corrige la grammaire, ajuste le ton, traduit ou resume. Instructions personnalisees supportees.', detail: 'gemma4 / llama3.2 / phi4 / personnalise' },
       { label: 'SORTIE', title: 'Insertion automatique', description: 'Le texte corrige est colle a la position du curseur dans n\'importe quelle app. Bloc-notes, VS Code, Chrome, Slack, partout.', detail: 'Presse-papiers + Ctrl+V / 100ms latence' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/ja.ts

@@ -47,7 +47,7 @@ export const ja: Translations = {
     steps: [
       { label: '入力', title: 'ホットキー押下', description: 'Right Alt（またはカスタムキー）を長押しで録音開始。ダブルタップでAIモード。トグルでハンズフリー。', detail: '長押し / トグル / ダブルプレス' },
       { label: '文字起こし', title: 'Whisper STT', description: 'ローカルfaster-whisperエンジンが16kHz PCMオーディオをリアルタイムでテキストに変換。GPU加速対応。', detail: 'faster-whisper / base ~ large-v3' },
-      { label: '処理', title: 'Ollama LLM', description: 'ローカルLLMが文法を校正し、トーンを調整し、翻訳または要約。カスタム指示に対応。', detail: 'qwen3 / llama3 / gemma3 / カスタム' },
+      { label: '処理', title: 'Ollama LLM', description: 'ローカルLLMが文法を校正し、トーンを調整し、翻訳または要約。カスタム指示に対応。', detail: 'gemma4 / llama3.2 / phi4 / カスタム' },
       { label: '出力', title: '自動挿入', description: '校正されたテキストが任意のアプリのカーソル位置に自動貼り付け。メモ帳、VS Code、Chrome、Slackなど。', detail: 'クリップボード + Ctrl+V / 100msレイテンシ' },
     ],
     panelTitle: 'D3RO Voice パイプライン',

site/src/i18n/locales/ko.ts

@@ -47,7 +47,7 @@ export const ko: Translations = {
     steps: [
       { label: '입력', title: '핫키 누르기', description: 'Right Alt(또는 설정한 키)를 길게 눌러 녹음을 시작. 더블탭으로 AI 모드. 토글로 핸즈프리.', detail: '길게 누르기 / 토글 / 더블프레스' },
       { label: '전사', title: 'Whisper STT', description: '로컬 faster-whisper 엔진이 16kHz PCM 오디오를 실시간으로 텍스트로 변환. GPU 가속 지원.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: '처리', title: 'Ollama LLM', description: '로컬 LLM이 문법을 다듬고, 톤을 조절하고, 번역하거나 요약합니다. 커스텀 명령어 지원.', detail: 'qwen3 / llama3 / gemma3 / 커스텀' },
+      { label: '처리', title: 'Ollama LLM', description: '로컬 LLM이 문법을 다듬고, 톤을 조절하고, 번역하거나 요약합니다. 커스텀 명령어 지원.', detail: 'gemma4 / llama3.2 / phi4 / 커스텀' },
       { label: '출력', title: '자동 삽입', description: '다듬어진 텍스트가 어떤 앱에서든 커서 위치에 자동 붙여넣기. 메모장, VS Code, Chrome, Slack, 어디서나.', detail: '클립보드 + Ctrl+V / 100ms 지연' },
     ],
     panelTitle: 'D3RO Voice 파이프라인',

site/src/i18n/locales/pt.ts

@@ -47,7 +47,7 @@ export const pt: Translations = {
     steps: [
       { label: 'ENTRADA', title: 'Pressione a tecla', description: 'Segure Right Alt (ou sua tecla personalizada) para gravar. Toque duplo para modo IA. Alternar para viva-voz.', detail: 'Segurar / Alternar / Toque duplo' },
       { label: 'TRANSCREVER', title: 'Whisper STT', description: 'Motor local faster-whisper converte audio PCM 16kHz em texto em tempo real. Aceleracao GPU suportada.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'PROCESSAR', title: 'Ollama LLM', description: 'LLM local corrige gramatica, ajusta tom, traduz ou resume. Instrucoes personalizadas suportadas.', detail: 'qwen3 / llama3 / gemma3 / personalizado' },
+      { label: 'PROCESSAR', title: 'Ollama LLM', description: 'LLM local corrige gramatica, ajusta tom, traduz ou resume. Instrucoes personalizadas suportadas.', detail: 'gemma4 / llama3.2 / phi4 / personalizado' },
       { label: 'SAIDA', title: 'Insercao automatica', description: 'Texto polido e colado na posicao do cursor em qualquer app. Bloco de notas, VS Code, Chrome, Slack, em qualquer lugar.', detail: 'Area de transferencia + Ctrl+V / 100ms latencia' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/ru.ts

@@ -47,7 +47,7 @@ export const ru: Translations = {
     steps: [
       { label: 'ВВОД', title: 'Нажмите горячую клавишу', description: 'Удерживайте Right Alt (или вашу клавишу) для записи. Двойное нажатие для ИИ-режима. Переключение для свободных рук.', detail: 'Удержание / Переключение / Двойное нажатие' },
       { label: 'ТРАНСКРИПЦИЯ', title: 'Whisper STT', description: 'Локальный движок faster-whisper преобразует 16кГц PCM аудио в текст в реальном времени. Поддержка GPU-ускорения.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'ОБРАБОТКА', title: 'Ollama LLM', description: 'Локальная LLM корректирует грамматику, настраивает тон, переводит или обобщает. Пользовательские инструкции поддерживаются.', detail: 'qwen3 / llama3 / gemma3 / пользовательский' },
+      { label: 'ОБРАБОТКА', title: 'Ollama LLM', description: 'Локальная LLM корректирует грамматику, настраивает тон, переводит или обобщает. Пользовательские инструкции поддерживаются.', detail: 'gemma4 / llama3.2 / phi4 / пользовательский' },
       { label: 'ВЫВОД', title: 'Автовставка', description: 'Отредактированный текст вставляется в позицию курсора в любом приложении. Блокнот, VS Code, Chrome, Slack, где угодно.', detail: 'Буфер обмена + Ctrl+V / 100мс задержка' },
     ],
     panelTitle: 'D3RO Voice Конвейер',

site/src/i18n/locales/vi.ts

@@ -47,7 +47,7 @@ export const vi: Translations = {
     steps: [
       { label: 'DAU VAO', title: 'Nhan phim nong', description: 'Giu Right Alt (hoac phim tuy chinh) de ghi am. Nhan doi cho che do AI. Chuyen doi cho che do ranh tay.', detail: 'Giu / Chuyen doi / Nhan doi' },
       { label: 'PHIEN AM', title: 'Whisper STT', description: 'Bo may faster-whisper cuc bo chuyen doi am thanh PCM 16kHz thanh van ban thoi gian thuc. Ho tro tang toc GPU.', detail: 'faster-whisper / base ~ large-v3' },
-      { label: 'XU LY', title: 'Ollama LLM', description: 'LLM cuc bo sua ngu phap, dieu chinh giong dieu, dich hoac tom tat. Ho tro chi dan tuy chinh.', detail: 'qwen3 / llama3 / gemma3 / tuy chinh' },
+      { label: 'XU LY', title: 'Ollama LLM', description: 'LLM cuc bo sua ngu phap, dieu chinh giong dieu, dich hoac tom tat. Ho tro chi dan tuy chinh.', detail: 'gemma4 / llama3.2 / phi4 / tuy chinh' },
       { label: 'DAU RA', title: 'Tu dong chen', description: 'Van ban da tinh chinh duoc dan vao vi tri con tro trong bat ky ung dung nao. Notepad, VS Code, Chrome, Slack, bat cu dau.', detail: 'Clipboard + Ctrl+V / 100ms do tre' },
     ],
     panelTitle: 'D3RO Voice Pipeline',

site/src/i18n/locales/zh.ts

@@ -47,7 +47,7 @@ export const zh: Translations = {
     steps: [
       { label: '输入', title: '按下快捷键', description: '长按 Right Alt（或自定义按键）开始录音。双击进入AI模式。切换免提模式。', detail: '长按 / 切换 / 双击' },
       { label: '转录', title: 'Whisper STT', description: '本地 faster-whisper 引擎将 16kHz PCM 音频实时转换为文本。支持GPU加速。', detail: 'faster-whisper / base ~ large-v3' },
-      { label: '处理', title: 'Ollama LLM', description: '本地 LLM 润色语法、调整语气、翻译或摘要。支持自定义指令。', detail: 'qwen3 / llama3 / gemma3 / 自定义' },
+      { label: '处理', title: 'Ollama LLM', description: '本地 LLM 润色语法、调整语气、翻译或摘要。支持自定义指令。', detail: 'gemma4 / llama3.2 / phi4 / 自定义' },
       { label: '输出', title: '自动插入', description: '润色后的文本自动粘贴到任何应用的光标位置。记事本、VS Code、Chrome、Slack，任何地方。', detail: '剪贴板 + Ctrl+V / 100ms延迟' },
     ],
     panelTitle: 'D3RO Voice 流程',

site/src/sections/HowItWorks.tsx

@@ -44,7 +44,7 @@ export function HowItWorks() {
                   <div className="flex items-center gap-2">
                     <Led color="green" size="sm" />
                     <span className="text-emerald-400/70">{t.pipeline.llmConnected}</span>
-                    <span className="text-neutral-600">qwen3:4b @ localhost:11434</span>
+                    <span className="text-neutral-600">gemma4:e4b @ localhost:11434</span>
                   </div>
                   <div className="h-px bg-white/[0.04] my-3" />
                   <div className="flex items-start gap-2">