Sync live code; move MQTT credentials out of source; untrack venv/

- mqtt_audio_bridge.py: MQTT_USER / MQTT_PASS are read from the environment,
  falling back to ~/.config/environment.d/10-secrets.conf (mode 600). The
  environment.d trap means a fresh SSH shell needs the file fallback.
- docker-compose.yml: pass MQTT_USER / MQTT_PASS through; keeps the live
  network_mode / Dockerfile topology.
- start_bridge.sh: launcher that sources the secrets file before compose up.
- app.py: commits the live initial_prompt that was uncommitted.
- .gitignore: add venv/ (was missing), __pycache__, *.bak*, *.wav.
- git rm -r --cached venv/: 13,413 tracked venv files removed from the index.

Credential note: MQTT_PASS was hardcoded in edaf914. origin/main was later
force-rewritten to redact it, but the value was exposed and must still be
rotated. Removing venv/ from the index does not shrink the repository: the
objects stay in history until a history rewrite and GC.

Co-Authored-By: Paperclip <noreply@paperclip.ing>

app.py

@@ -24,7 +24,17 @@ def transcribe_audio():
 
     try:
         # Now we can safely use temp_file_path
-        segments, info = model.transcribe(temp_file_path, beam_size=5, language="en")
+        initial_prompt = (
+            "Jervis, feed the fish, turn on the lights, turn off the lights, "
+            "shopping list, add to the shopping list, calendar, weather, time, "
+            "remind me, heater, watering the plants, play music, stop"
+        )
+        segments, info = model.transcribe(
+            temp_file_path,
+            beam_size=5,
+            language="en",
+            initial_prompt=initial_prompt,
+        )
         transcript = " ".join([segment.text for segment in segments])
         print(f"Transcript: {transcript}")
         return jsonify({"transcript": transcript})