PRD: Gemma 4 brain integration into Vector AI system

[chatde]

Overview

The Vector AI Brain at /Volumes/AI-Models/vector/ already has a production-grade autonomous system:

• Priority-based BehaviorArbiter (SAFETY > ONBOARD > AI > IDLE)
• Persistent MemoryBank with Jaccard scene similarity
• LearningEngine for autonomous environment discovery
• MissionController (SecurityPatrol, ObjectCensus, RoomCheck, PresenceWatch)
• SpatialMemory for room mapping
• VoiceHandler for voice commands

Current AI stack: LLaVA (vision) + Llama 3.1 (personality) — two separate models, no memory context injection.

Goal

Replace both models with Gemma 4 (single call) that:

1. Sees through Vector's camera (multimodal vision)
2. Gets injected with full MemoryBank context — every prior observation, known objects, room history
3. Returns unified nav decision + personality reaction + memory note in one JSON response
4. Plugs transparently into the existing BehaviorArbiter + MemoryBank + MissionController

Architectural decisions

• Swap mechanism: GEMMA4=1 env var in rover_brain.py — zero changes to existing code paths
• Model: gemma4:e4b via Ollama REST at http://localhost:11434
• Context injection: MemoryBank.get_recent(5) + known_objects + SpatialMemory summary prepended to every prompt
• Output contract: JSON with keys: scene, objects, nav_decision, danger_level, reaction, memory_note
• Patcher: patch_all_models() monkey-patches mission_controller._ask_vision and learning_engine LLaVA calls
• Fallback: if Gemma 4 unavailable, falls back to original AIBrain (LLaVA)

Delivered so far

• gemma4_brain.py — Gemma4Brain class + patch_all_models() + standalone CLI
• rover_brain.py — GEMMA4=1 swap point added
• gemmi_processor.py — Gemma 4 replaces LLaVA as Ollama fallback
• gemma4_companion.py — lightweight companion loop using existing SDK config

What remains

See child issues for implementation slices.