Fix port collision: scope PORT env to llama.cpp subprocess

entrypoint.sh

@@ -11,7 +11,10 @@ CTX="${PHANTOM_GRID_LLAMACPP_CONTEXT_LENGTH:-8192}"
 APP_PORT="${PORT:-7860}"
 
 echo "[entrypoint] Starting llama.cpp server on :${LLAMA_PORT} with ${MODEL_PATH}"
-python -m llama_cpp.server \
+# llama_cpp.server's pydantic Settings reads HOST/PORT from env and they
+# override the CLI flags — so on HF Spaces (PORT=7860) the server would bind
+# 7860 and collide with Gradio. Scope HOST/PORT to just this subprocess.
+HOST=127.0.0.1 PORT="${LLAMA_PORT}" python -m llama_cpp.server \
     --model "${MODEL_PATH}" \
     --host 127.0.0.1 \
     --port "${LLAMA_PORT}" \