fix(cilium,spire): set resource requests to promote critical agents out of BestEffort QoS

k8s/bases/infrastructure/controllers/cilium/helm-release.yaml

@@ -79,6 +79,10 @@ spec:
           enabled: false
     operator:
       replicas: ${cilium_replicas:=2}
+      resources:
+        requests:
+          cpu: 100m
+          memory: 256Mi
       podDisruptionBudget:
         enabled: true
         minAvailable: 1
@@ -94,6 +98,24 @@ spec:
     ipam:
       mode: kubernetes
     kubeProxyReplacement: true
+    # ------------------------------------------------------------------
+    # Resource requests for the agent DaemonSet and the standalone
+    # cilium-envoy DaemonSet. These promote the pods out of BestEffort
+    # QoS so they survive node memory pressure; an OOMKilled cilium-agent
+    # leaves BPF state degraded and the node loses ClusterIP routing
+    # (observed cascading into ~13 workload crash-loops on prod-worker-2,
+    # 2026-05-28). Limits intentionally unset — Cilium recommends against
+    # capping the agent (https://docs.cilium.io/en/stable/operations/performance/).
+    # ------------------------------------------------------------------
+    resources:
+      requests:
+        cpu: 200m
+        memory: 512Mi
+    envoy:
+      resources:
+        requests:
+          cpu: 50m
+          memory: 128Mi
     # Transparent WireGuard encryption for all pod-to-pod and node-to-node
     # traffic.  KubeSpan (Talos-layer WireGuard between nodes) is not
     # enabled in this cluster, so without this setting inter-node pod
@@ -118,8 +140,23 @@ spec:
           install:
             namespace: kube-system
             existingNamespace: true
+            # Resource requests promote spire-server and spire-agent pods
+            # out of BestEffort QoS. cilium-agent's SPIRE Delegate API
+            # client relies on the per-node spire-agent admin socket — if
+            # the agent is evicted/OOMKilled the cilium-agent on that
+            # node stays stuck retrying "SPIRE admin socket does not
+            # exist" and ClusterIP routing degrades alongside it.
+            agent:
+              resources:
+                requests:
+                  cpu: 50m
+                  memory: 128Mi
             # TODO: Remove workaround when SPIRE no longer fails to start (https://github.com/cilium/cilium/issues/40533)
             server:
+              resources:
+                requests:
+                  cpu: 50m
+                  memory: 128Mi
               initContainers:
               - command:
                 - /bin/sh