Full app audit: 24 fixes across backend, frontend, infra, content, social

Critical fixes:
- Fix hangup-during-respond crash (null caller guard)
- Fix double-click caller race condition
- Stem recorder: non-daemon thread, disk error handling, 30s flush timeout
- Frontend startCall() error handling

High priority:
- Devon: filter tool errors from speech, shorter monitor prompt, 30s interval
- TTS ghost message fix (add to history after TTS, not before)
- Expand banned phrase list (12 new phrases)
- Increase returning callers from 1 to 2 per session
- Platform-tailored social posts with staggered scheduling
- YouTube dynamic tags from episode content
- Social post retry logic (2 attempts, 5s delay)
- Frontend: error handling on all raw fetch calls

Medium:
- stem_recorder null check race (local var capture in audio.py)
- Reactive shape directive expanded
- REACT TO LUKE moved higher in caller prompt
- Devon tenure updated ("few weeks" not "first day")
- D shortcut Escape to unfocus
- Volume slider debounced (150ms)
- Settings modal widened to 550px
- Backup script (daily MariaDB dump + data/ rsync to NAS)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/services/llm.py

@@ -236,7 +236,7 @@ class LLMService:
                 try:
                     result = await tool_executor(tool_name, arguments)
                 except Exception as e:
-                    result = f"Error: {e}"
+                    result = f"Tool unavailable — could not complete {tool_name} right now."
                     print(f"[LLM-Tools] Tool {tool_name} failed: {e}")
 
                 all_tool_calls.append({