clean: archive legacy stt/tts/llm services; update ARCHITECTURE.md + README.md to current stack (REST gpt-4o-transcribe, nano, sage, Honcho, incremental TTS, white noise)

Per PLAN item 6. Legacy files moved to archive/legacy-pipeline/.

README.md

@@ -34,7 +34,7 @@ cd ~/Projects/ai-body-double
 cp .env.example .env
 # Edit .env with your API keys:
 #   OPENAI_API_KEY=sk-...
-#   DEEPSEEK_API_KEY=sk-...
+#   HONCHO_API_KEY=...
 ```
 
 ### 3. Run
@@ -56,17 +56,19 @@ kira.hobokenchicken.com {
 ## Architecture
 
 ```
-Browser ──WebSocket──▶ Backend (FastAPI)
-  │                        │
-  ├─ Mic audio ──────────▶ ├─ Whisper API (STT)
-  │                        ├─ DeepSeek (LLM)
-  │ ◀── TTS audio ──────── ├─ OpenAI TTS
-  │                        │
-  ├─ YouTube embed (lo-fi) │
-  ├─ Timer / Notes / Cats  │
-  └─ Animated avatar       │
+Browser (React + Live2D + WhiteNoise + Notes) ──WebSocket──▶ Backend (FastAPI)
+  │                                                              │
+  ├─ Mic (MediaRecorder full webm on stop) ────────────────────▶ ├─ gpt-4o-transcribe (REST STT, emits delta)
+  │                                                              ├─ gpt-5.4-nano + Honcho memory context
+  │ ◀── Incremental Opus chunks (play on arrival) ─────────────── ├─ OpenAI TTS streaming (sage)
+  │                                                              │
+  ├─ YouTube lofi + Web Audio noise (independent)                │
+  ├─ Timer / Notes / Pets / Wardrobe / Scenes                    │
+  └─ Live2D (or fallback) avatar                                 │
 ```
 
+Note: Realtime WebSocket STT (gpt-realtime-whisper) was attempted for true streaming but blocked by model access — current REST path is stable and cheap.
+
 ## Live2D Model Setup
 
 Kira currently uses a CSS/SVG animated placeholder avatar. To add a Live2D model: