Voice command upgrade: answer_question() with camera + memory context

[chatde]

Parent PRD

#1

What to build

Upgrade VoiceHandler to route voice queries through Gemma4Brain.answer_question() instead of the old Llama 3.1 text model. Voice queries now get full memory context + optional camera frame, so Vector gives richer, personalized answers.

Acceptance criteria

• Asking Vector 'What do you see?' uses Gemma4Brain.describe_view(pil_image) — not LLaVA
• Asking 'What have you seen today?' references MemoryBank recent observations
• Asking 'What is that?' with a pointing gesture analyzes the current camera frame
• Response is spoken via say_text() with TTS chunking
• Voice queries work when GEMMA4=1 env var is set
• If camera is unavailable, answer_question() still works text-only

Blocked by

• Blocked by rover_brain.py GEMMA4=1 integration + live E2E test #5 (E2E test must pass first)

User stories addressed

• Voice commands get intelligent, memory-aware answers from Gemma 4