GPUStack With vLLM in Production: Pinning KV Cache Before the OOM Finds You

Integrating GPUStack with vLLM: pinned KV cache limits, a hardened serve block, and an aiohttp TTFT/RSS soak test that catches pool decay early.