rover_brain.py GEMMA4=1 integration + live E2E test

[chatde]

Parent PRD

#1

What to build

Wire Gemma4Brain into the full RoverBrain startup path via the GEMMA4=1 env var, and validate the complete system runs end-to-end with a real Vector robot: face detection still works (onboard), BehaviorArbiter arbitrates correctly, Gemma 4 provides nav decisions, memory persists across restart.

Acceptance criteria

• GEMMA4=1 python3 main.py boots without errors
• Logger shows: 🧠 Gemma 4 brain ACTIVE — LLaVA + Llama3.1 replaced
• BehaviorArbiter.translate_ai_decision() correctly handles Gemma4Brain output
• Vector's onboard face detection still triggers (ONBOARD priority, not replaced)
• Gemma 4 nav decision (AI priority) is overridden by safety events correctly
• After 2 minutes of roaming, MemoryBank has at least 3 new Gemma 4 observations
• Restarting with GEMMA4=1 loads prior MemoryBank — Gemma 4 references past observations
• GEMMA4=0 python3 main.py (or without flag) still uses original AIBrain (LLaVA)

Blocked by

• Blocked by Gemma4Brain: core process_frame() — unified vision + nav JSON #2 (core), Memory context injection — feed MemoryBank + SpatialMemory into every Gemma 4 call #3 (memory), patch_all_models() — upgrade MissionController + LearningEngine to Gemma 4 #4 (patcher)

User stories addressed

• Full Vector AI system runs on Gemma 4 with a single env var
• Existing face detection, cliff safety, and mission features all preserved