feat: Smart Queue Consumer implementation draft + architecture review

- SMART_QUEUE_IMPLEMENTATION.md: Complete implementation draft (1572 lines)
  with 10 quick-win fixes and full smart queue consumer rewrite
- ARCHITECTURE_REVIEW.md: 26-issue audit with prioritized findings
- Verified all 3 GPUs live: amdpve (73% util), llmgpu (idle), ocu_llm (idle)
- Redis 7.4.9 confirmed streams support
- GPU sidecar metrics verified on all hosts

Key fixes:
- QW-1: Dockerfile path mismatch (Dockerfile.queue -> queue-service/Dockerfile)
- QW-2: Nginx fallback only on ALL-GPU failure (not single GPU)
- QW-3: Container names fixed to Docker service names
- QW-4: Redis host default fixed (192.168.68.7 -> redis)
- QW-5: Dependency version pinning
- QW-7-10: Health checks, restart policy, Gunicorn, single-process collector

Smart queue features:
- Redis Streams + consumer groups
- GPU-aware load balancing via sidecar metrics
- Per-GPU circuit breakers with half-open recovery
- Adaptive backpressure (0-30 normal, 30-40 warn, 40-50 503, >50 open)
- Dead letter queue with retry endpoint
- Job ID tracking and /status/<job_id> API

gpu-router-docker.conf

@@ -24,7 +24,7 @@ upstream queue_service {
 
 upstream dashboard_service {
     ## Harness dashboard (Docker container)
-    server dashboard:3001;
+    server syslog-harness-dashboard-1:3001;
 }
 
 upstream gpu_dashboard_pool {