Switch to bridge networking with fixed container port 80 and served name llm

- run.sh: bridge + -p ${HOST_PORT:-24396}:80, health check uses host port
- entrypoint.sh: PORT defaults to 80, SERVED_NAME defaults to llm
  detected port/served name become log hints only
- verify.sh: read host port from port bindings, expect model name llm
- docker-compose.yml: ports mapping instead of network_mode host
- README: update tables, commands and rationale

entrypoint.sh

@@ -34,18 +34,22 @@ if [[ "${AUTO_DETECT}" == "1" ]]; then
   MODEL_KIND="${MODEL_KIND:-${DETECTED_MODEL_KIND}}"
   TP="${TP:-${DETECTED_TP}}"
   PP="${PP:-${DETECTED_PP}}"
-  PORT="${PORT:-${DETECTED_PORT}}"
-  SERVED_NAME="${SERVED_NAME:-${DETECTED_SERVED_NAME}}"
   MODEL_DESCRIPTION="${DETECTED_DESCRIPTION:-}"
+  # 检测出的 port / served_name 仅用于日志提示，不再决定容器内监听端口和模型名：
+  # 容器内端口固定 80，模型名固定 llm（见 AGENTS.md 5.2 / 5.5）。
+  DETECTED_HOST_PORT_HINT="${DETECTED_PORT:-}"
+  DETECTED_NAME_HINT="${DETECTED_SERVED_NAME:-}"
 else
   MODEL_KIND="${MODEL_KIND:-unknown}"
+  DETECTED_HOST_PORT_HINT=""
+  DETECTED_NAME_HINT=""
 fi
 
 HOST="${HOST:-0.0.0.0}"
 TP="${TP:-8}"
 PP="${PP:-1}"
-PORT="${PORT:-8091}"
-SERVED_NAME="${SERVED_NAME:-llama4}"
+PORT="${PORT:-80}"
+SERVED_NAME="${SERVED_NAME:-llm}"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-1}"
 GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
@@ -78,6 +82,10 @@ echo "  extra      : ${EXTRA_ARGS}"
 if [[ -n "${MODEL_DESCRIPTION}" ]]; then
   echo "  detected   : ${MODEL_DESCRIPTION}"
 fi
+if [[ -n "${DETECTED_HOST_PORT_HINT}" ]]; then
+  echo "  hint       : 变体默认宿主端口 ${DETECTED_HOST_PORT_HINT}，模型名曾用 ${DETECTED_NAME_HINT}"
+  echo "               （容器内固定 80 / served name 固定 llm，实际端口由 docker run -p 决定）"
+fi
 echo "============================================================"
 
 # shellcheck disable=SC2086