diff --git a/product/infra/helm/evolith-agent-runtime/templates/service.yaml b/product/infra/helm/evolith-agent-runtime/templates/service.yaml
index 5addf4daa..95252975d 100644
--- a/product/infra/helm/evolith-agent-runtime/templates/service.yaml
+++ b/product/infra/helm/evolith-agent-runtime/templates/service.yaml
@@ -11,5 +11,11 @@ spec:
       targetPort: http
       protocol: TCP
       name: http
+      {{- if and (eq .Values.service.type "NodePort") .Values.service.nodePort }}
+      # PINNED. An unpinned NodePort is allocated at random from 30000-32767 and
+      # changes on reinstall, so anything configured against it breaks — which is
+      # the failure a stable cross-cluster address exists to prevent.
+      nodePort: {{ .Values.service.nodePort }}
+      {{- end }}
   selector:
     app: {{ include "evolith-agent-runtime.name" . }}
diff --git a/product/infra/helm/evolith-agent-runtime/values-local.yaml b/product/infra/helm/evolith-agent-runtime/values-local.yaml
index 6d0c6959e..d1dfc8983 100644
--- a/product/infra/helm/evolith-agent-runtime/values-local.yaml
+++ b/product/infra/helm/evolith-agent-runtime/values-local.yaml
@@ -1,5 +1,14 @@
-# Local Kubernetes (Docker Desktop) overrides for evolith-agent-runtime.
-# Reach the service with: kubectl -n evolith-local port-forward svc/<release>-evolith-agent-runtime 8082:80
+# Local Kubernetes (Docker Desktop / kind) overrides for evolith-agent-runtime.
+#
+# Two ways to reach it, and they are not interchangeable:
+#
+#   FROM ANOTHER CLUSTER (the Tracker) — http://<cluster>-control-plane:30082
+#   FROM THIS MACHINE, ad hoc          — kubectl -n evolith-local port-forward svc/<release>-evolith-agent-runtime 8082:80
+#                                        or http://localhost:30082 on a kind cluster built with core-cluster.yaml
+#
+# The cross-cluster one is why `service.type` is NodePort here: a port-forward is
+# a host process that dies on every rollout, so it cannot be what a separate
+# cluster is configured against.
 image:
   repository: evolith-agent-runtime
   tag: local2
@@ -26,3 +35,7 @@ podDisruptionBudget:
 
 networkPolicy:
   enabled: false
+
+service:
+  type: NodePort
+  nodePort: 30082
diff --git a/product/infra/helm/evolith-agent-runtime/values.yaml b/product/infra/helm/evolith-agent-runtime/values.yaml
index 620b09ddd..b907a6a65 100644
--- a/product/infra/helm/evolith-agent-runtime/values.yaml
+++ b/product/infra/helm/evolith-agent-runtime/values.yaml
@@ -18,6 +18,12 @@ service:
   type: ClusterIP
   port: 80
   targetPort: 3000
+  # Only read when `type: NodePort`. Empty means "let Kubernetes allocate one",
+  # which is fine for a throwaway and wrong for anything another cluster is
+  # configured against — see values-local.yaml and kind/core-cluster.yaml.
+  # ClusterIP stays the default: a node port is a LOCAL cross-cluster
+  # affordance, never a production one.
+  nodePort: ""
 
 ingressRoute:
   enabled: true
diff --git a/product/infra/helm/evolith-core-api/templates/service.yaml b/product/infra/helm/evolith-core-api/templates/service.yaml
index dbd58b440..a146fb2b6 100644
--- a/product/infra/helm/evolith-core-api/templates/service.yaml
+++ b/product/infra/helm/evolith-core-api/templates/service.yaml
@@ -11,5 +11,12 @@ spec:
       targetPort: http
       protocol: TCP
       name: http
+      {{- if and (eq .Values.service.type "NodePort") .Values.service.nodePort }}
+      # PINNED on purpose. An unpinned NodePort is allocated at random from
+      # 30000-32767, so it changes on reinstall — and a cross-cluster consumer
+      # configured against it would break every time the chart is reinstalled,
+      # which is the failure `extraPortMappings` exists to prevent.
+      nodePort: {{ .Values.service.nodePort }}
+      {{- end }}
   selector:
     app: {{ include "evolith-core-api.name" . }}
diff --git a/product/infra/helm/evolith-core-api/values-local.yaml b/product/infra/helm/evolith-core-api/values-local.yaml
index 0b436abb9..a5dde4e57 100644
--- a/product/infra/helm/evolith-core-api/values-local.yaml
+++ b/product/infra/helm/evolith-core-api/values-local.yaml
@@ -1,6 +1,16 @@
-# Local Kubernetes (Docker Desktop) overrides for evolith-core-api.
+# Local Kubernetes (Docker Desktop / kind) overrides for evolith-core-api.
 # Use the locally-built image, skip Traefik/IngressRoute and cluster-only addons.
-# Reach the service with: kubectl -n evolith-local port-forward svc/<release>-evolith-core-api 8080:80
+#
+# Two ways to reach it, and they are not interchangeable:
+#
+#   FROM THIS MACHINE, ad hoc — kubectl -n evolith-local port-forward svc/<release>-evolith-core-api 8080:80
+#   FROM ANOTHER CLUSTER (the Tracker) — http://host.docker.internal:30080
+#
+# The second is why `service.type` is NodePort here. A port-forward is a process
+# on the host that dies on every rollout of the Core, so it cannot be what a
+# separate cluster is configured against; the node port is fixed at cluster
+# creation and survives. Requires the cluster to have been created with
+# `product/infra/kind/core-cluster.yaml` — the mapping cannot be added later.
 image:
   repository: evolith-core-api
   tag: local2
@@ -8,6 +18,10 @@ image:
 
 replicaCount: 1
 
+service:
+  type: NodePort
+  nodePort: 30080
+
 ingressRoute:
   enabled: false
 
diff --git a/product/infra/helm/evolith-core-api/values.yaml b/product/infra/helm/evolith-core-api/values.yaml
index 7b7c69ba2..6971d30bf 100644
--- a/product/infra/helm/evolith-core-api/values.yaml
+++ b/product/infra/helm/evolith-core-api/values.yaml
@@ -22,6 +22,12 @@ service:
   type: ClusterIP
   port: 80
   targetPort: 3000
+  # Only read when `type: NodePort`. Empty means "let Kubernetes allocate one",
+  # which is fine for a throwaway and wrong for anything another cluster is
+  # configured against — see values-local.yaml and kind/core-cluster.yaml.
+  # ClusterIP stays the default: exposing the Core on a node port is a LOCAL
+  # cross-cluster affordance, never a production one.
+  nodePort: ""
 
 ingressRoute:
   enabled: true
diff --git a/product/infra/helm/evolith-mcp/templates/service.yaml b/product/infra/helm/evolith-mcp/templates/service.yaml
index acfcfb111..8d1d8633c 100644
--- a/product/infra/helm/evolith-mcp/templates/service.yaml
+++ b/product/infra/helm/evolith-mcp/templates/service.yaml
@@ -11,6 +11,12 @@ spec:
       targetPort: http
       protocol: TCP
       name: http
+      {{- if and (eq .Values.service.type "NodePort") .Values.service.nodePort }}
+      # PINNED. An unpinned NodePort is allocated at random from 30000-32767 and
+      # changes on reinstall, so anything configured against it breaks — which is
+      # the failure a stable cross-cluster address exists to prevent.
+      nodePort: {{ .Values.service.nodePort }}
+      {{- end }}
     {{- if .Values.opa.enabled }}
     # GT-551: expose the OPA sidecar's HTTP port so Prometheus can reach its
     # /metrics endpoint. Without this the sidecar is pod-local only and
diff --git a/product/infra/helm/evolith-mcp/values-local.yaml b/product/infra/helm/evolith-mcp/values-local.yaml
index 9cab0c99b..e2e46f9ff 100644
--- a/product/infra/helm/evolith-mcp/values-local.yaml
+++ b/product/infra/helm/evolith-mcp/values-local.yaml
@@ -1,7 +1,16 @@
-# Local Kubernetes (Docker Desktop) overrides for evolith-mcp.
+# Local Kubernetes (Docker Desktop / kind) overrides for evolith-mcp.
 # The OPA sidecar is DISABLED locally (it fetches signed bundles from an in-cluster
-# MinIO that does not exist on Docker Desktop). Reach the service with:
-#   kubectl -n evolith-local port-forward svc/<release>-evolith-mcp 8081:80
+# MinIO that does not exist on Docker Desktop).
+#
+# Two ways to reach it, and they are not interchangeable:
+#
+#   FROM ANOTHER CLUSTER (the Tracker) — http://<cluster>-control-plane:30081
+#   FROM THIS MACHINE, ad hoc          — kubectl -n evolith-local port-forward svc/<release>-evolith-mcp 8081:80
+#                                        or http://localhost:30081 on a kind cluster built with core-cluster.yaml
+#
+# The cross-cluster one is why `service.type` is NodePort here: a port-forward is
+# a host process that dies on every rollout, so it cannot be what a separate
+# cluster is configured against.
 image:
   repository: evolith-mcp
   tag: local2
@@ -23,3 +32,7 @@ podDisruptionBudget:
 
 networkPolicy:
   enabled: false
+
+service:
+  type: NodePort
+  nodePort: 30081
diff --git a/product/infra/helm/evolith-mcp/values.yaml b/product/infra/helm/evolith-mcp/values.yaml
index 159c41001..f9d6749a1 100644
--- a/product/infra/helm/evolith-mcp/values.yaml
+++ b/product/infra/helm/evolith-mcp/values.yaml
@@ -19,6 +19,12 @@ service:
   port: 80
   # The real mcp-server listens on 3000 (packages/mcp-server/Dockerfile).
   targetPort: 3000
+  # Only read when `type: NodePort`. Empty means "let Kubernetes allocate one",
+  # which is fine for a throwaway and wrong for anything another cluster is
+  # configured against — see values-local.yaml and kind/core-cluster.yaml.
+  # ClusterIP stays the default: a node port is a LOCAL cross-cluster
+  # affordance, never a production one.
+  nodePort: ""
 
 # App container environment. The mcp-server requires EVOLITH_API_KEY in production
 # (injected from a secret below); set EVOLITH_MCP_ALLOW_NO_AUTH=true to relax.
diff --git a/product/infra/helm/local-test.sh b/product/infra/helm/local-test.sh
index dba722601..ce6e41509 100644
--- a/product/infra/helm/local-test.sh
+++ b/product/infra/helm/local-test.sh
@@ -8,6 +8,7 @@
 #   bash product/infra/helm/local-test.sh kind-apps-up # kind + build + load + install apps
 #   bash product/infra/helm/local-test.sh kind-up      # kind + build + load + infra + apps
 #   bash product/infra/helm/local-test.sh smoke        # port-forward + curl
+#   bash product/infra/helm/local-test.sh url          # cross-cluster URL, and proof it answers
 #   bash product/infra/helm/local-test.sh kind-down    # uninstall + delete kind cluster
 #
 # Env:
@@ -27,16 +28,143 @@ CLUSTER="${KIND_CLUSTER:-evolith}"   # dedicated kind cluster name (kind-<name>
 
 IMAGES=(evolith-core-api:"$IMAGE_TAG" evolith-mcp:"$IMAGE_TAG" evolith-agent-runtime:"$IMAGE_TAG")
 
+KIND_CONFIG="$ROOT/product/infra/kind/core-cluster.yaml"
+CORE_NODE_PORT=30080
+# One entry per interface: "<label> <nodePort>". The Core is not one address —
+# REST, MCP and the agent runtime are three surfaces at parity, and a check that
+# covered only the first would report a parity it never exercised.
+SURFACES=("core-api 30080" "mcp 30081" "agent-runtime 30082")
+# The address ANOTHER cluster uses to reach this Core.
+#
+# Every kind cluster joins the same `kind` Docker network, so the consumer's pod
+# reaches this cluster's NODE CONTAINER by name — Docker's embedded DNS resolves
+# it and the NodePort is listening there. This path does NOT use the host port
+# mapping, and it must not: that one is bound to 127.0.0.1 and is therefore
+# unreachable from the Docker bridge.
+#
+# NOT `host.docker.internal`: that fails with `Could not resolve host` inside a
+# kind pod, which resolves through CoreDNS in the node and never sees the Docker
+# Desktop entry. Measured, after writing it the wrong way first.
+CORE_CROSS_CLUSTER_URL="http://$CLUSTER-control-plane:$CORE_NODE_PORT"
+
 kind_create() {
   if kind get clusters 2>/dev/null | grep -qx "$CLUSTER"; then
     echo "==> kind cluster '$CLUSTER' already exists"
+    # A cluster created before core-cluster.yaml has no host port, and NOTHING
+    # can add one to a running cluster — port mappings are fixed at creation.
+    # Saying so is the whole point: otherwise the Tracker is configured against
+    # a port that was never bound, and the failure surfaces later as a refused
+    # connection with no obvious cause.
+    if ! docker inspect "$CLUSTER-control-plane" --format '{{json .HostConfig.PortBindings}}' 2>/dev/null | grep -q "$CORE_NODE_PORT/tcp"; then
+      echo "!!  This cluster has NO host mapping for $CORE_NODE_PORT — it predates product/infra/kind/core-cluster.yaml."
+      echo "!!  CROSS-CLUSTER access is unaffected ($CORE_CROSS_CLUSTER_URL goes over the shared 'kind'"
+      echo "!!  Docker network, not through this mapping). What you lose is reaching it from THIS machine"
+      echo "!!  at http://localhost:$CORE_NODE_PORT — a browser, a curl, the CLI."
+      echo "!!  To gain that, recreate the cluster (DESTRUCTIVE — everything in it is lost):"
+      echo "!!      bash $0 kind-down && bash $0 kind-up"
+    fi
   else
-    echo "==> Creating kind cluster '$CLUSTER'"
-    kind create cluster --name "$CLUSTER"
+    echo "==> Creating kind cluster '$CLUSTER' with the host port mapping ($KIND_CONFIG)"
+    kind create cluster --name "$CLUSTER" --config "$KIND_CONFIG"
   fi
   kubectl config use-context "kind-$CLUSTER"
 }
 
+# Print every surface's address and PROVE each one, rather than asserting them.
+#
+# Three surfaces x two paths, and they fail INDEPENDENTLY: the host mapping is
+# fixed at cluster creation, the cross-cluster path depends only on the shared
+# Docker network, and each service has its own pinned port. One aggregate check
+# would let a green core-api vouch for an MCP nobody reached.
+url() {
+  local rc=0 peer=""
+
+  echo "cross-cluster base (configure the Tracker with this): http://$CLUSTER-control-plane:<port>"
+  echo "from this machine:                                    http://localhost:<port>"
+  for s in "${SURFACES[@]}"; do
+    set -- $s
+    printf "  %-14s %s\n" "$1" "$2"
+  done
+
+  echo
+  echo "== GET /health from THIS MACHINE =="
+  for s in "${SURFACES[@]}"; do
+    set -- $s
+    if curl -fsS --max-time 5 "http://localhost:$2/health" >/dev/null 2>&1; then
+      printf "   %-14s OK   (localhost:%s)\n" "$1" "$2"
+    else
+      printf "!! %-14s FAIL (localhost:%s)\n" "$1" "$2"
+      rc=1
+    fi
+  done
+  if [ "$rc" -ne 0 ]; then
+    echo "!! In order: does the cluster carry the mappings (docker inspect $CLUSTER-control-plane),"
+    echo "!! are the services NodePort on those ports (kubectl -n $NS get svc), are the pods ready?"
+  fi
+
+  echo
+  echo "== GET /health FROM ANOTHER CLUSTER =="
+  # Any other RUNNING kind cluster will do. `kind get clusters` also lists
+  # stopped ones, and the first version of this picked one: the probe failed,
+  # blamed the Docker network, and the real cause was a node container that had
+  # been dead for two days. A peer that cannot answer is not evidence.
+  for c in $(kind get clusters 2>/dev/null | grep -vx "$CLUSTER"); do
+    if docker inspect "$c-control-plane" --format '{{.State.Running}}' 2>/dev/null | grep -qx true; then
+      peer="$c"; break
+    fi
+  done
+  if [ -z "$peer" ]; then
+    echo "   skipped: no OTHER RUNNING kind cluster to call from."
+    echo "   This is NOT a pass — the cross-cluster path is simply unexercised."
+    return $rc
+  fi
+
+  for s in "${SURFACES[@]}"; do
+    set -- $s
+    # Declared separately: under `set -u`, a single `local a=1 b=$a` does not see
+    # `a` yet and dies with "unbound variable".
+    local label="$1" port="$2"
+    local pod="xcluster-probe-$label-$$"
+    local out=""
+    # NOT `kubectl run -i --rm`. That attaches to the pod AFTER creating it, so a
+    # container which has already exited loses its stdout — and the result moves
+    # between runs for reasons that have nothing to do with the thing under test.
+    # Observed directly: two consecutive runs disagreed about WHICH surface was
+    # reachable. Create, wait, read the logs, delete.
+    #
+    # stderr is kept throughout: discarding it is how a failed image pull reads
+    # as "the Docker network is broken".
+    kubectl --context "kind-$peer" run "$pod" --restart=Never \
+      --image=curlimages/curl:8.10.1 --command -- \
+      curl -fsS --max-time 10 "http://$CLUSTER-control-plane:$port/health" >/dev/null 2>&1
+    kubectl --context "kind-$peer" wait --for=jsonpath='{.status.phase}'=Succeeded \
+      "pod/$pod" --timeout=60s >/dev/null 2>&1 \
+      || kubectl --context "kind-$peer" wait --for=jsonpath='{.status.phase}'=Failed \
+           "pod/$pod" --timeout=5s >/dev/null 2>&1 || true
+    out="$(kubectl --context "kind-$peer" logs "$pod" 2>&1)$(kubectl --context "kind-$peer" get pod "$pod" -o jsonpath='{.status.containerStatuses[*].state.waiting.reason}' 2>&1)"
+    kubectl --context "kind-$peer" delete pod "$pod" --now >/dev/null 2>&1 || true
+
+    # Case-INSENSITIVE and whitespace-tolerant, because the three surfaces do not
+    # agree on the shape: core-api answers in the ADR-0073 envelope with
+    # `"status":"OK"` nested under `data`, mcp and agent-runtime return a bare
+    # object with `"status":"ok"`. The first version matched only the former, so
+    # two healthy services were reported unreachable and the printed remedy
+    # blamed the network. A probe that cannot read the answer manufactures a defect.
+    if printf '%s' "$out" | grep -qiE '"status"[[:space:]]*:[[:space:]]*"ok"'; then
+      printf "   %-14s OK   (from cluster '%s')\n" "$label" "$peer"
+    elif printf '%s' "$out" | grep -qiE "ErrImagePull|ImagePullBackOff|failed to pull"; then
+      printf "!! %-14s INCONCLUSIVE — the probe pod could not start (curlimages/curl did not pull).\n" "$label"
+      echo   "!! That says nothing about the cross-cluster path. Check registry connectivity."
+      rc=1
+    else
+      printf "!! %-14s FAIL at http://%s-control-plane:%s\n" "$label" "$CLUSTER" "$port"
+      printf '%s\n' "$out" | tail -2 | sed 's/^/!!   /'
+      rc=1
+    fi
+  done
+  return $rc
+}
+
 kind_load() {
   echo "==> Loading images into kind cluster '$CLUSTER'"
   for img in "${IMAGES[@]}"; do
@@ -202,9 +330,10 @@ case "${1:-kind-up}" in
          kind_create; build; kind_load; secrets; install
          echo "==> Done. Run: bash $0 smoke" ;;
   smoke) smoke ;;
+  url)   url ;;
   down)  down ;;
   kind-down)
          down
          kind delete cluster --name "$CLUSTER" ;;
-  *) echo "usage: $0 {kind-apps-up|kind-up|apps-up|up|build|smoke|down|kind-down}"; exit 1 ;;
+  *) echo "usage: $0 {kind-apps-up|kind-up|apps-up|up|build|smoke|url|down|kind-down}"; exit 1 ;;
 esac
diff --git a/product/infra/kind/core-cluster.yaml b/product/infra/kind/core-cluster.yaml
new file mode 100644
index 000000000..37dc05fa5
--- /dev/null
+++ b/product/infra/kind/core-cluster.yaml
@@ -0,0 +1,86 @@
+# kind cluster config for the Core, with ONE stable host port.
+#
+# ## Why this file exists
+#
+# The Core and the Tracker are separate products in separate repositories and
+# they talk over HTTP. Validating them in ONE cluster would give them in-cluster
+# DNS and hide the class of defect that only appears across a real network
+# boundary: a malformed base URL, a missing auth header, a timeout, CORS. Two
+# clusters is the honest topology, and it is already the de-facto one —
+# `kind/tracker-web.local.yaml` deploys into a separate `evolith-tracker`
+# cluster.
+#
+# What was missing is the LINK. Until now everything in `local-test.sh` was
+# reached with `kubectl port-forward`, which is right for a smoke test inside one
+# cluster and wrong as the bridge between two:
+#
+#   - it is a process on the host, so pods in the OTHER cluster cannot depend on
+#     it — there is nothing for the Tracker's config to point at that survives;
+#   - it DIES ON EVERY ROLLOUT of the target. Redeploy the Core and the Tracker
+#     stops seeing it, and the symptom looks like a Tracker defect.
+#
+# A pinned NodePort gives a fixed address that survives rollouts, restarts and
+# re-installs, so a consumer is configured once and stays configured.
+#
+# ## There are TWO addresses, and they are not interchangeable — measured, not assumed
+#
+#   FROM ANOTHER CLUSTER  http://<cluster>-control-plane:30080
+#   FROM THIS MACHINE     http://localhost:30080
+#
+# The cross-cluster one does NOT go through the host mapping below. Every kind
+# cluster joins the same `kind` Docker network, so the consumer's pod reaches
+# this cluster's node container directly, by container name — Docker's embedded
+# DNS resolves it and the NodePort is listening there. Verified on 2026-08-03: a
+# pod in a second kind cluster got `{"status":"OK"}` from this Core, by name and
+# by IP, and again after `rollout restart`.
+#
+# **`host.docker.internal` does not work from inside a kind pod** — it fails with
+# `Could not resolve host`. It is a Docker Desktop convenience for containers the
+# daemon runs directly; a pod resolves through CoreDNS inside the node and never
+# sees that entry. This note exists because that URL was the first thing written
+# here and it was wrong.
+#
+# The host mapping below is for the OTHER consumer — a browser, a curl, a CLI on
+# the laptop. It is bound to 127.0.0.1, which is also why it cannot serve the
+# cross-cluster case: a loopback binding is unreachable from the Docker bridge.
+# Keeping the scopes separate is deliberate: the machine-local port stays
+# machine-local, and cluster-to-cluster traffic never leaves the Docker network.
+#
+# ## The constraint worth knowing before you run it
+#
+# Port mappings are fixed AT CLUSTER CREATION. They cannot be added to a running
+# kind cluster — `docker inspect evolith-cluster-control-plane` on a cluster made
+# without this file shows only `6443/tcp` bound, and no `kubectl` command adds
+# one. Adopting this means DELETING and RECREATING the cluster, which is why
+# `local-test.sh` does not do it silently: `kind_create` tells you when the
+# existing cluster predates this config and leaves the decision to you.
+#
+# ## Port choice
+#
+# 30080 is inside Kubernetes' default NodePort range (30000-32767), so it can be
+# pinned in values without `--service-node-port-range`. It is bound on 127.0.0.1
+# and not on 0.0.0.0 on purpose: this is a development cluster carrying a
+# dev-auth API key, and it has no business being reachable from the network the
+# laptop happens to be on.
+kind: Cluster
+apiVersion: kind.x-k8s.io/v1alpha4
+nodes:
+  - role: control-plane
+    extraPortMappings:
+      # One per INTERFACE, not one for the Core. "Evolith Core and its
+      # interfaces" means REST, MCP and the agent runtime, and a consumer that
+      # can reach only the first has not exercised the product — the surfaces
+      # are meant to be at parity, so a validation that covers one of three
+      # would report parity it never tested.
+      - containerPort: 30080   # core-api  — REST
+        hostPort: 30080
+        listenAddress: "127.0.0.1"
+        protocol: TCP
+      - containerPort: 30081   # mcp       — MCP
+        hostPort: 30081
+        listenAddress: "127.0.0.1"
+        protocol: TCP
+      - containerPort: 30082   # agent-runtime
+        hostPort: 30082
+        listenAddress: "127.0.0.1"
+        protocol: TCP