Card: separate the two RAM mechanisms in the setup table + Setup A - host K/V is the decode-time stream pin (int8/fp16/q4_0 only); the RAM-backed context between sessions is the conversation cache + prompt cache, which k8v4 keeps fully
Card: add Deployment setups (Setup A deep-context/ours vs Setup B concurrent, KV-in-VRAM vs host K/V trade-offs by workload); correct the KV section - host K/V is OFF by design under k8v4 (221k pcie=0.0%)
