Qwen3-Embedding-8B via vLLM pooling — 4096-dim embeddings for CCA code intelligence.