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

docker-compose.yml

@@ -9,12 +9,14 @@ services:
     # 镜像本身不设置启动命令，启动方式全部由这里决定。
     entrypoint: ["/usr/local/bin/llama4-entrypoint.sh"]
     working_dir: /opt/llama4
-    expose:
-      - "8091"
-      - "8092"
 
-    # vLLM 多卡 NCCL / 端口直通
-    network_mode: host
+    # 容器内固定监听 80，宿主机端口由 HOST_PORT 决定（默认 24396）。
+    expose:
+      - "80"
+    ports:
+      - "${HOST_PORT:-24396}:80"
+
+    # bridge 网络 + 端口映射；ipc host 供 vLLM/PyTorch 多卡共享内存。
     ipc: host
 
     # Iluvatar 运行时要求
@@ -33,12 +35,14 @@ services:
       # 默认暴露全部 16 张卡；entrypoint 会根据模型选择实际 TP*PP。
       CUDA_VISIBLE_DEVICES: "${GPUS:-0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}"
 
+      # 容器内端口固定 80，模型名固定 llm（AGENTS.md 5.2 / 5.5）。
+      PORT: "80"
+      SERVED_NAME: "${SERVED_NAME:-llm}"
+
       # 留空时由 entrypoint 根据 config.json 自动判断。
       # 需要手动覆盖时直接设置同名环境变量即可。
       TP: "${TP:-}"
       PP: "${PP:-}"
-      PORT: "${PORT:-}"
-      SERVED_NAME: "${SERVED_NAME:-}"
 
       MAX_MODEL_LEN: "${MAX_MODEL_LEN:-4096}"
       MAX_NUM_SEQS: "${MAX_NUM_SEQS:-1}"