Tarquin: StackChan robot butler on Muse Spark

# Otter-chan / Tarquin

Custom firmware and server for the **M5Stack StackChan** (CoreS3, ESP32-S3) that turns it into
*Tarquin*: a Victorian butler with a dry wit, a British male voice, a camera that follows you, and a
bridge to your Grok bots.

```
 StackChan (firmware/)  ── WiFi/WebSocket ──►  Tarquin server (server/, Go)  ──►  Muse Spark 1.3 (Meta Model API)
   mic → VAD → PCM                                ├─ parakeet-server (STT, ggml)   ◄──  Grok bots (REST, bearer)
   speaker ◄ TTS PCM                              └─ otter-vox (Audio8 TTS, ggml)
   camera → motion + face assist
   servos, LEDs, face on screen
```

No Python at runtime. Speech runs on CPU via ggml (parakeet.cpp for STT, otter-vox for TTS).

## Features

- **Side button**: click = open mic until you stop talking; click again = close mic; double-click = hard
  mute (wake word off); hold = sleep / wake. The bottom button is the hardware reset line and cannot be
  intercepted by software (it reboots), so sleep lives on the side button hold and on `/api/sleep`.
- **Wake word "tarquin"** (hands-free, on-device): a microWakeWord streaming model trained here
  (`tools/wakeword/`) runs on the ESP32-S3 via tflite-micro, ~1 % CPU, no network. Falls back to shipping
  short clips for server-side matching if no model is compiled in.
- **Butler**: Muse Spark 1.3 (contributor tier by default) with tools: message bots, read bot status,
  take and look at a photo (multimodal), timers, notes/memory, gestures, lights, volume, sleep, tracking.
- **TTS**: piper by default (native, ~25× real time on CPU, British male voices `en_GB-alan-medium` /
  `en_GB-northern_english_male-medium`), or `OTTER_TTS_ENGINE=vox` for the Audio8 zero-shot "tarquin" clone
  (needs a strong GPU to be quick). Sentence streaming starts speech while the model is still writing.
- **Screen**: animated robot face (blinks, gaze follows you, mouth syncs to speech, expressions from the
  LLM) plus live cards for each Grok bot's activity (coding cursor, browsing globe, thinking dots…).
- **Tracking**: camera motion centroid on-device + server face detection (pure Go, pigo) → head follows you;
  returns home when you leave.
- **Fun**: head-touch tap = happy nod, swipe = volume, hold = hearts; shake it = "I am not a cocktail";
  screen tap = wink, hold = IP/battery; idle fidgets; LED patterns per state; dance gesture; timers spoken aloud.