fix(agentx): keep Dynamo discovery alive during cold start

Apply ETCD_LEASE_TTL=7200 globally in both GB200 AgentX recipes so the frontend and all workers retain discovery leases through the two-hour health window.

中文：在两个 GB200 AgentX 配方中全局设置 ETCD_LEASE_TTL=7200，确保前端及所有 worker 在最长两小时的健康检查窗口内维持 Dynamo 服务发现租约。

Author: ivanium

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-agentic.yaml

@@ -14,6 +14,11 @@ dynamo:
 
 setup_script: vllm-container-deps.sh
 
+environment:
+  # The frontend shares Grace CPU capacity with the long TP8 cold start. Keep
+  # Dynamo discovery alive until the outer two-hour health check expires.
+  ETCD_LEASE_TTL: "7200"
+
 slurm:
   time_limit: "8:00:00"
 
@@ -59,7 +64,6 @@ backend:
       enable_offload: false
   aggregated_environment:
     DYN_REQUEST_PLANE: "tcp"
-    ETCD_LEASE_TTL: "600"
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
     TILELANG_CLEANUP_TEMP_FILES: "1"

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p1d-dep8-dep16-agentic.yaml

@@ -14,6 +14,11 @@ dynamo:
 
 setup_script: vllm-container-deps.sh
 
+environment:
+  # Apply the lease to the frontend as well as every prefill/decode worker so
+  # discovery survives the long CPU-saturated cold start.
+  ETCD_LEASE_TTL: "7200"
+
 slurm:
   time_limit: "8:00:00"
 
@@ -65,14 +70,6 @@ backend:
       mode: "embedded"
       enable_offload: false
   prefill_environment:
-    # dynamo's primary etcd lease defaults to a 10s TTL. During the ~380s
-    # weight load the process is CPU-saturated and the Rust keep-alive task
-    # cannot renew in time, so etcd revokes the lease and worker registration
-    # later fails with "requested lease not found" (workers never go ready ->
-    # health check times out at 2h). 600s covers the weight load + cudagraph
-    # stalls; longer TTL only delays dead-worker cleanup, which the health
-    # check / fail-fast monitor handle independently.
-    ETCD_LEASE_TTL: "600"
     DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-agentx"
     TORCH_SYMMMEM: "NVSHMEM"
     VLLM_USE_NCCL_SYMM_MEM: "1"
@@ -101,9 +98,6 @@ backend:
     # MC_STORE_CLIENT_METRIC_INTERVAL: "5"
     # MC_TE_METRIC: "0"
   decode_environment:
-    # See prefill_environment: raise dynamo's primary etcd lease TTL from the
-    # 10s default so the lease survives the CPU-saturated cold start.
-    ETCD_LEASE_TTL: "600"
     DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-agentx"
     TORCH_SYMMMEM: "NVSHMEM"
     VLLM_USE_NCCL_SYMM_MEM: "1"