feat(llama-swap): pin to GPU 0, reserving it for large-model hot-loads

ana-ml2's Ada->Blackwell swap (2x96GB) frees GPU 0 entirely. Pin llama-swap
to GPU 0 via NVIDIA_VISIBLE_DEVICES so on-demand large-model hot-loads land
there, off GPU 1 where the always-on vLLM services (granite + embed/rerank/
reward) live. Closes the long-standing 'pin llama-swap to GPU 0' item.

stacks/llama-swap/compose.yaml

@@ -29,6 +29,12 @@ services:
     environment:
       - HF_HOME=/hfcache
       - HF_HUB_CACHE=/hfcache/hub
+      # Pin to GPU 0 — the reserved card for on-demand large-model hot-loads.
+      # The always-on vLLM services (granite + embed/rerank/reward) own GPU 1;
+      # keeping llama-swap off GPU 1 stops a hot-loaded model from contending
+      # with them. llama.cpp then sees only GPU 0 (cuda:0), so --n-gpu-layers
+      # 999 loads there with no per-model device targeting needed.
+      - NVIDIA_VISIBLE_DEVICES=${LLAMA_SWAP_GPU:-0}
     healthcheck:
       test: ["CMD-SHELL", "curl -fsS http://localhost:8080/ >/dev/null || exit 1"]
       interval: 30s