refactor(ray.sub): drop NETWORK_INIT_CMDS — MC_TCP_BIND_ADDRESS suffices

The NETWORK_INIT_CMDS block (pkill avahi-autoipd / ifconfig usb0 down /
ip addr flush + a 2-second relaunch loop) was a workaround for an
outdated diagnosis in data-plane-bench/DEBUG_TQ_BACKENDS.md (Issue 1):
"MC_TCP_BIND_ADDRESS controls server_name (registration) but NOT the
RPC listener bind address."

Re-reading current Mooncake main (commit fast-forwarded today):

  - mooncake-transfer-engine/src/transfer_engine_impl.cpp:159-170
    If MC_TCP_BIND_ADDRESS is set, it goes directly into
    desc.ip_or_host_name, which is the address registered via
    addRpcMetaEntry — i.e. the address peers receive from the
    metadata service. This was added by PR #226 (caef1ef, merged
    2025-04-10) and IS in the pinned wheel 0.3.10.post2 (bumped
    2026-04-22).

  - mooncake-transfer-engine/src/transfer_metadata_plugin.cpp:1292
    The TCP listener binds INADDR_ANY and accepts on all interfaces.
    Bind itself was never the bug — the announce was.

So per-process MC_TCP_BIND_ADDRESS in TQDataPlaneClient.__init__
(unchanged in this commit, runs on every process) gives Mooncake the
routable announce address and peer connections work cross-node
without OS-level interface stripping.

The pkill+sleep loop fought a symptom (avahi-autoipd respawning the
APIPA address). With the announce now correct regardless of usb0,
that fight is unnecessary. Removing the block also makes ray.sub a
no-op for non-mooncake_cpu backends (simple, mooncake_rdma) — they
were paying the host-process-kill cost for no reason.

If multi-node smoke regresses with peers connecting to 169.254.x,
revert this commit only — (A) codec/adapter cleanup stays.

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

Author: ZhiyuLi-Nvidia

ray.sub

@@ -205,40 +205,10 @@ head_node_ip=${ip_addresses_array[0]}
 
 ip_head=$head_node_ip:$PORT
 
-# Network init for Mooncake-cpu (TQ data-plane backend "mooncake_cpu").
-# Mooncake's transfer_metadata_plugin.cpp:1127 calls getifaddrs() and binds
-# the RPC listener to the FIRST interface with an IP — usually usb0
-# (169.254.3.1 link-local APIPA), which is unreachable cross-node. The fix
-# kills avahi-autoipd (the daemon that re-assigns 169.254.3.1), tells
-# NetworkManager to stop managing usb0, flushes the address, and runs a
-# 2 s relaunch loop as a failsafe. Lifted from data-plane-bench/ray.sub
-# (proven at 32-node and 48-node scales). Belt-and-braces: ifconfig +
-# ip variants both attempted because the container set varies.
-# Without this, mooncake_cpu fails with metadata 404s and a
-# MemcpyWorkerPool segfault during the first kv_batch_put.
-# See research/data_plane_mooncake_status.md.
-NETWORK_INIT_CMDS='# Kill avahi-autoipd for usb0: it is the daemon that re-assigns 169.254.3.1.
-pkill avahi-autoipd 2>/dev/null || true
-if [ -f /run/avahi-autoipd.usb0.pid ]; then kill $(cat /run/avahi-autoipd.usb0.pid) 2>/dev/null || true; fi
-nmcli device set usb0 managed no 2>/dev/null || true
-ifconfig usb0 0.0.0.0 2>/dev/null || true
-ifconfig usb0 down 2>/dev/null || true
-ip link set usb0 down 2>/dev/null || true
-ip addr flush dev usb0 2>/dev/null || true
-{ while :; do
-    pkill avahi-autoipd 2>/dev/null || true
-    ifconfig usb0 0.0.0.0 2>/dev/null || true
-    ifconfig usb0 down 2>/dev/null || true
-    ip link set usb0 down 2>/dev/null || true
-    ip addr flush dev usb0 2>/dev/null || true
-    sleep 2
-  done; } &'
-
 # First we start the head of the ray cluster on one of the physical nodes
 # Give the head node actual resources to make it schedulable
 
 head_cmd=$(cat <<EOF
-$NETWORK_INIT_CMDS
 # Touch a file to indicate that the head node has started
 # Overlapping srun commands will check this file to determine if we can overlap a container command
 touch $LOG_DIR/STARTED_RAY_HEAD
@@ -347,7 +317,6 @@ for ((i = 1; i < SLURM_JOB_NUM_NODES; i++)); do
   node_i=${nodes_array[$i]}
 
   worker_cmd=$(cat <<EOF
-$NETWORK_INIT_CMDS
 env
 
 exit-dramatically() {