feat: 배포 파이프라인 — Ollama/sidecar 번들 + NSIS 자동 VC++ + GitLab CI + 온보딩 모달

- sidecar 슬림화: torch/pyannote 제거, ctranslate2 GPU 감지, /diarize 삭제
- Ollama 번들: resources/ollama/에 포터블 바이너리 배치, LocalLLMService 1순위 탐색
- installer.nsh: VC++ 재배포 x64 자동 다운로드(aka.ms 경유) + 사일런트 설치
- electron-builder: extraResources에 ollama 추가, nsis.include로 installer.nsh 연결
- scripts: download-ollama.ps1/sh 신규
- LLM.PULL_MODEL IPC 핸들러 + LocalLLMService.pullModel() 구현 (api/pull 스트리밍)
- 온보딩 모달: gemma4:e4b 미설치 감지 시 자동 표시, 진행률 UI, i18n(ko/en) 키 추가
- .gitlab-ci.yml: Windows 러너에서 sidecar/sox/ollama 준비 후 NSIS 패키징, 태그 시 Release 자동 생성

apps/desktop/src/main/ipc/llm-handlers.ts

@@ -24,6 +24,21 @@ export function registerLLMHandlers(): void {
     safeSendToRenderer(IPC_CHANNELS.LLM.STATUS_CHANGED, { status: llm.getStatus() })
   })
 
+  // Pull 진행률 → 렌더러
+  llm.on('pull-progress', (payload: unknown) => {
+    safeSendToRenderer(IPC_CHANNELS.LLM.PULL_PROGRESS, payload)
+  })
+
+  ipcMain.handle(IPC_CHANNELS.LLM.PULL_MODEL, async (_event, params: { modelId: string }) => {
+    try {
+      await getLocalLLMService().pullModel(params.modelId)
+      return ipcSuccess(undefined)
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err)
+      return ipcError(ErrorCode.LLMServerUnreachable, `Pull 실패: ${msg}`)
+    }
+  })
+
   // Phase 3.2: VoiceModeService의 premium-llm-fallback 이벤트를 렌더러로 전달
   getVoiceModeService().on('premium-llm-fallback', (payload: { reason: string }) => {
     safeSendToRenderer(IPC_CHANNELS.LLM.PREMIUM_FALLBACK, payload)

apps/desktop/src/main/services/LocalLLMService.ts

@@ -11,6 +11,7 @@ import { configGet } from './ConfigService'
 import { D3ROError, ErrorCode } from '@d3ro/core/errors'
 import type { LLMStatus, LLMModel, LLMAction, LLMConnectionState } from '@d3ro/core/types'
 import { resolveSystemPrompt } from './llm-prompts'
+import { getBundledOllamaPath } from '../utils/paths'
 
 const logger = getLogger('LocalLLMService')
 
@@ -167,6 +168,12 @@ class LocalLLMService extends EventEmitter {
   private async _findOllamaBinary(): Promise<string | null> {
     const candidates: string[] = []
 
+    // 1순위: 설치 파일에 번들된 ollama
+    const bundled = getBundledOllamaPath()
+    if (bundled) {
+      candidates.push(bundled)
+    }
+
     if (process.platform === 'win32') {
       const localAppData = process.env.LOCALAPPDATA
       if (localAppData) {
@@ -482,6 +489,77 @@ class LocalLLMService extends EventEmitter {
     }
   }
 
+  /**
+   * Ollama /api/pull — 모델 다운로드. 진행률을 EventEmitter로 방출.
+   * 스트리밍 JSON 라인을 파싱해 각 chunk마다 'pull-progress' 이벤트 emit.
+   * 완료 시 resolve, 에러 시 reject.
+   *
+   * @param modelId 예: 'gemma4:e4b'
+   */
+  async pullModel(modelId: string): Promise<void> {
+    const serverUrl = configGet('ollamaServerUrl')
+    logger.info(`Pull 시작: ${modelId}`)
+
+    const response = await fetch(`${serverUrl}/api/pull`, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ model: modelId, stream: true })
+    })
+
+    if (!response.ok || !response.body) {
+      throw new D3ROError(
+        ErrorCode.LLMServerUnreachable,
+        `Pull 실패: HTTP ${response.status}`
+      )
+    }
+
+    const reader = response.body.getReader()
+    const decoder = new TextDecoder()
+    let buffer = ''
+
+    for (;;) {
+      const { done, value } = await reader.read()
+      if (done) break
+
+      buffer += decoder.decode(value, { stream: true })
+      const lines = buffer.split('\n')
+      buffer = lines.pop() ?? ''
+
+      for (const line of lines) {
+        const trimmed = line.trim()
+        if (trimmed.length === 0) continue
+        try {
+          const chunk = JSON.parse(trimmed) as {
+            status: string
+            digest?: string
+            total?: number
+            completed?: number
+            error?: string
+          }
+          if (chunk.error) {
+            throw new D3ROError(ErrorCode.LLMServerUnreachable, chunk.error)
+          }
+          this.emit('pull-progress', {
+            modelId,
+            status: chunk.status,
+            digest: chunk.digest ?? null,
+            total: chunk.total ?? 0,
+            completed: chunk.completed ?? 0,
+            percent:
+              chunk.total && chunk.total > 0
+                ? Math.min(100, Math.floor(((chunk.completed ?? 0) / chunk.total) * 100))
+                : 0
+          })
+        } catch (err) {
+          if (err instanceof D3ROError) throw err
+          // JSON 파싱 실패한 부분 라인은 스킵
+        }
+      }
+    }
+
+    logger.info(`Pull 완료: ${modelId}`)
+  }
+
   getStatus(): LLMStatus {
     const connectionState: LLMConnectionState = this._available
       ? this._state === LLMState.Generating

apps/desktop/src/main/utils/paths.ts

@@ -88,6 +88,20 @@ export function getSidecarCommand(): { command: string; args: string[] } {
   return { command: pythonCmd, args: [sidecarPath] }
 }
 
+/**
+ * 번들된 Ollama 실행 파일 경로. 존재하지 않으면 null을 반환해 시스템 설치본 탐색으로 폴백.
+ * - Windows: ollama.exe
+ * - macOS/Linux: ollama
+ */
+export function getBundledOllamaPath(): string | null {
+  const ollamaBin = `ollama${EXE_SUFFIX}`
+  const bundled = isPackaged()
+    ? path.join(process.resourcesPath, 'ollama', ollamaBin)
+    : path.join(app.getAppPath(), 'resources', 'ollama', ollamaBin)
+
+  return existsSync(bundled) ? bundled : null
+}
+
 /**
  * 효과음 파일 경로.
  */