llama.cpp issue #26529: with the RPC backend active, llama-server aborts inside rpc_buffer_get_tensor the first time a slot prompt becomes eligible for prompt-cache save

Reporter ClosedLadder, build 11924d4c1, GNU 13.3.0, Linux aarch64, 3x NVIDIA GB10 (DGX Spark) with CUDA backend, coordinator + 2 ggml-rpc-server workers (one over ConnectX-7 P2P, one over gigabit LAN), Qwen3-235B-A22B-Instruct-2507 Q6_K split GGUF, 4 shards, --tensor-split 45,40,15 -ngl 999 --split-mode layer. Deterministic: server answers a short first request then dies on the first request large enough to trigger a cache save. Single reporter, reproducible on demand.