feat: guided light meditation + Edge TTS neural voice audio

- Add continuous guided light-body-scan meditation (16 segments, 97% coverage)
- Edge TTS API endpoint with Xiaoxiao neural voice at -40% rate
- Pre-generated chime WAVs and guidance MP3s in public/audio/
- Mute toggle in practice page, sequenced non-overlapping audio
- Fix middleware to serve /audio and /api/audio/tts without auth
- All LLM output localized to Chinese; fix blank home/settings pages
- DeepSeek API compat: json_object, increased max_tokens
- Analytics, cron reminders, onboarding flow fixes

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/lib/engines/summary-generator.ts

@@ -58,7 +58,9 @@ Voice rules:
 - Be honest. If it was a light week with only 1-2 sessions, acknowledge that without judgment.
 - Never diagnose, never claim medical benefits, never use self-help clichés.
 - Write in second person ("you").
-- Keep it to 3-4 paragraphs, 200-300 words total.`;
+- Keep it to 3-4 paragraphs, 200-300 words total.
+
+IMPORTANT: The user interface is Chinese-only. Write the entire summary in Simplified Chinese (简体中文). Use "你" (not "you").`;
 
 export async function generateWeeklySummary(
   userId: string,
@@ -123,7 +125,7 @@ export async function generateWeeklySummary(
         },
       ],
       temperature: 0.7,
-      max_tokens: 500,
+      max_tokens: 2000,
     });
 
     const summary = response.choices[0]?.message?.content?.trim();