test(e2e): capture_node writes unprefixed node/nf and can clobber its caller's loop variable #3675

capture_node() in hack/e2e-capture-dataplane.sh assigns node=$1 and nf="$OUT/node-$node.txt" without any prefix or scoping, while its only caller binds node in the loop it runs from:

while IFS='|' read -r ns pod podip node _ready _phase; do
  ...
  capture_node "$node"
  ...
  agent=$(pod_on_node "$CILIUM_NS" k8s-app=cilium "$node")

Harmless today, because the caller passes the same $node it just read, so the write lands on the value already there. It stops being harmless the moment someone calls capture_node with a different node inside that loop: the assignment silently rewrites the loop variable, and the rest of that iteration follows it. Downstream of the call, the rest of the iteration reads $node in its pod lookups (k8s-app=cilium, app=ovs, app=kube-ovn-cni) and in the headers it writes. The cni lookup is the one that hurts: ip route get and conntrack run through that pod, so a swapped node yields another machine's routing and conntrack state, filed under the right pod's name. Nothing fails; the output just describes the wrong node, and reads as if it did not.

The newer code in the same file uses _cpd_ and _cpr_ prefixes for exactly this reason, so this is one of the last spots on the old convention.

Fix: prefix both variables the way the rest of the file does. The script targets POSIX sh, so local is not the tool here.

Pre-existing, not introduced by anything currently in flight.