Add fully local conversational AI pipeline for Reachy Mini

Local STT (Qwen3-ASR), VLM (Gemma 4 26B-A4B), and TTS (Spark-TTS) running
on Apple Silicon via MLX, with bracket-tag action system for nod, shake,
wiggle, dance, photo, and pre-recorded emotions.

record_voice.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Record a voice reference via the Reachy Mini's microphone.
+# Usage: ./record_voice.sh [seconds] [output.wav]
+set -e
+
+DURATION="${1:-12}"
+OUTPUT="${2:-voice_ref.wav}"
+
+echo "Recording ${DURATION}s via Reachy Mini's microphone to ${OUTPUT}"
+echo "Stand/sit near the robot. Starting in 3 seconds..."
+
+sshpass -p 'root' ssh -o StrictHostKeyChecking=no pollen@reachy-mini.local "/venvs/mini_daemon/bin/python -c \"
+import time, wave, numpy as np
+from reachy_mini import ReachyMini
+
+with ReachyMini() as mini:
+    mini.media.start_recording()
+    time.sleep(1.5)  # warm-up + countdown buffer
+    print('GO — speak now', flush=True)
+
+    sr = mini.media.get_input_audio_samplerate()
+    chunks = []
+    start = time.time()
+    while time.time() - start < ${DURATION}:
+        sample = mini.media.get_audio_sample()
+        if sample is not None:
+            mono = sample.mean(axis=1) if sample.ndim == 2 else sample
+            chunks.append(mono)
+        else:
+            time.sleep(0.01)
+    mini.media.stop_recording()
+
+    audio = np.concatenate(chunks).astype(np.float32)
+    peak = float(np.abs(audio).max())
+    print(f'Captured {len(audio)/sr:.1f}s @ {sr}Hz, peak={peak:.3f}', flush=True)
+
+    # Normalize + convert to int16
+    if peak > 0:
+        audio = audio / peak * 0.9
+    pcm = (audio * 32767).astype(np.int16)
+
+    with wave.open('/tmp/voice_ref.wav', 'wb') as f:
+        f.setnchannels(1); f.setsampwidth(2); f.setframerate(sr)
+        f.writeframes(pcm.tobytes())
+\""
+
+sshpass -p 'root' scp pollen@reachy-mini.local:/tmp/voice_ref.wav "${OUTPUT}"
+echo "Saved ${OUTPUT}"