fix(vllm): raise default VLLM_MAX_MODEL_LEN 2048 → 65536

Compose default was 2048 tokens — far too low for RAG. 65536 fits all pipeline calls including doc-level contextual retrieval (~16K tokens).