diff --git a/hack/e2e-apps/talos-image-cache.sh b/hack/e2e-apps/talos-image-cache.sh
index 38d8e21ff..70760a735 100644
--- a/hack/e2e-apps/talos-image-cache.sh
+++ b/hack/e2e-apps/talos-image-cache.sh
@@ -2,33 +2,142 @@
 # e2e CRs.
 #
 # Prefer the in-sandbox caching mirror (hack/e2e-talos-image-cache.yaml, deployed
-# by hack/e2e-install-cozystack.bats) when it is Available, otherwise fall back to
-# the chart default (the public factory.talos.dev) by emitting nothing. The
-# mirror rides out the public factory's flaky/range-less egress that otherwise
-# stalls worker DataVolume imports past the 12-minute node-join deadline — see
-# cozystack/cozystack#3231. Falling back to the default means the mirror can only
-# help, never make CI worse.
+# by hack/e2e-install-cozystack.bats) when a worker's CDI importer can actually
+# reach it, otherwise fall back to the chart default (the public factory.talos.dev)
+# by emitting nothing. The mirror rides out the public factory's flaky/range-less
+# egress that otherwise stalls worker DataVolume imports past the 12-minute
+# node-join deadline — see cozystack/cozystack#3231. Falling back to the default
+# means the mirror can only help, never make CI worse.
+#
+# Reachability is not assumed. Tenant namespaces run under a default-deny Cilium
+# egress installed by the tenant chart: a worker's importer may reach the `world`
+# entity (hence the public factory) and the tenant's own tree, but not an
+# arbitrary kube-system Service. Without an explicit allow, the importer resolves
+# talos-image-cache.kube-system.svc yet its TCP connect to the ClusterIP is
+# silently dropped and the disk import never starts. This helper installs that
+# allow (the CiliumClusterwideNetworkPolicy shipped in the mirror manifest) and
+# then verifies the path end to end from a Pod that faces the exact same egress
+# rules as a real importer, so the mirror is used only when it genuinely works.
 
 TALOS_IMAGE_CACHE_SVC_URL="${TALOS_IMAGE_CACHE_SVC_URL:-http://talos-image-cache.kube-system.svc}"
 _TALOS_IMAGE_FACTORY_DECISION_FILE="${_TALOS_IMAGE_FACTORY_DECISION_FILE:-/tmp/e2e-talos-image-factory-url}"
+TALOS_IMAGE_CACHE_MANIFEST="${TALOS_IMAGE_CACHE_MANIFEST:-hack/e2e-talos-image-cache.yaml}"
+# Tenant namespace the kubernetes-* CRs are created in (run-kubernetes.sh and the
+# oidc bats all target tenant-test); the reachability probe runs here so it inherits
+# that namespace's default-deny egress, exactly as a worker importer does.
+TALOS_IMAGE_CACHE_PROBE_NS="${TALOS_IMAGE_CACHE_PROBE_NS:-tenant-test}"
+TALOS_IMAGE_CACHE_PROBE_POD="${TALOS_IMAGE_CACHE_PROBE_POD:-talos-image-cache-probe}"
 
-# _talos_image_cache_serves_ranges: succeed iff the mirror answers a byte-range
-# request with 206 Partial Content. Runs curl inside the serve pod against
-# localhost, range-probing the seeded image file it finds under /data, so the
-# check works without the sandbox being able to reach the ClusterIP.
-_talos_image_cache_serves_ranges() {
-  local pod code
+# _apply_talos_image_cache_egress_policy: install the CiliumClusterwideNetworkPolicy
+# that lets tenant CDI importer Pods egress to the mirror. It lives in the mirror
+# manifest but is applied here, not in hack/e2e-install-cozystack.bats, because that
+# manifest is applied before Cozystack (and thus before Cilium's CRDs) exists.
+# Idempotent; best-effort — a failure just leaves the probe below to fall back.
+_apply_talos_image_cache_egress_policy() {
+  command -v yq >/dev/null 2>&1 || return 1
+  yq 'select(.kind == "CiliumClusterwideNetworkPolicy")' "$TALOS_IMAGE_CACHE_MANIFEST" 2>/dev/null \
+    | kubectl apply -f - >/dev/null 2>&1
+}
+
+# _talos_image_cache_probe_overrides: emit the kubectl --overrides JSON for the
+# probe Pod. $1 is the container image, $2 the shell command it runs. The single
+# container replaces the one `kubectl run` generates (merge patch semantics), and
+# the security context satisfies the tenant namespace's restricted PSA.
+#
+# Pure string builder, kept separate so it can be unit-tested: an edit that
+# unbalances a quote here would emit invalid JSON, every probe would fail, and CI
+# would silently fall back to the public factory instead of erroring.
+_talos_image_cache_probe_overrides() {
+  printf '{"spec":{"securityContext":{"runAsNonRoot":true,"runAsUser":1000,"seccompProfile":{"type":"RuntimeDefault"}},"containers":[{"name":"c","image":"%s","command":["sh","-ec","%s"],"securityContext":{"allowPrivilegeEscalation":false,"capabilities":{"drop":["ALL"]}}}]}}' \
+    "$1" "$2"
+}
+
+# _talos_image_cache_probe_succeeded: succeed iff the probe output ($1) carries the
+# strict 206 marker.
+#
+# Gate strictly on 206. The old localhost check kept a reachable-but-range-less
+# mirror and only warned; here any non-206 — unreachable (code=000) or a plain 200 —
+# falls back to the public factory instead. serve.py always answers a valid range
+# with 206, so the 200 case is unreachable in practice; the strict gate keeps the
+# "use the mirror only when it is fully working" rule simple.
+#
+# Pure predicate, kept separate so it can be unit-tested: this single comparison is
+# what decides whether tenant workers are pointed at the mirror at all.
+_talos_image_cache_probe_succeeded() {
+  printf '%s' "$1" | grep -q 'code=206'
+}
+
+# _talos_image_cache_reachable_from_tenant: succeed iff a Pod that faces the exact
+# same tenant egress restrictions and network path as a worker's CDI importer can
+# reach the mirror Service ClusterIP with a byte-range (206) response.
+#
+# This is the load-bearing gate: a green result guarantees a real importer will
+# reach the mirror. It runs a throwaway Pod in the tenant namespace, labelled
+# cdi.kubevirt.io=importer (so the tenant egress default-deny plus the allow
+# policy above apply to it just as they do to a real importer), and has it curl
+# the ClusterIP Service for the seeded image with a Range header — exercising DNS,
+# the egress policy, ClusterIP translation, and range support together, not merely
+# that the server answers on localhost.
+#
+# cdi.kubevirt.io=importer is CDI's own importer-pod label. The probe and the
+# egress policy's endpointSelector must both use it; if a future CDI renames it,
+# they would still agree with each other but no longer with reality, and the probe
+# would pass while real importers stayed blocked. It is correct as of the CDI
+# version this platform ships (packages/system/kubevirt-cdi-operator).
+_talos_image_cache_reachable_from_tenant() {
+  local pod rel img curl_cmd overrides out
   pod=$(kubectl -n kube-system get pod \
     -l app.kubernetes.io/name=talos-image-cache \
     --field-selector=status.phase=Running \
     -o jsonpath='{.items[0].metadata.name}' 2>/dev/null) || return 1
   [ -n "$pod" ] || return 1
-  code=$(kubectl -n kube-system exec "$pod" -c serve -- sh -ec '
-    f=$(find /data -name "*.raw.xz" 2>/dev/null | head -n1)
-    [ -n "$f" ] || exit 1
-    curl -s -o /dev/null -w "%{http_code}" -r 0-0 "http://127.0.0.1:8080/${f#/data/}"
-  ' 2>/dev/null) || return 1
-  [ "$code" = "206" ]
+  # Discover the seeded file (its presence also confirms the seed finished) and
+  # probe that exact path, so the probe requests what a worker actually requests.
+  rel=$(kubectl -n kube-system exec "$pod" -c serve -- sh -ec \
+    'f=$(find /data -name "*.raw.xz" 2>/dev/null | head -n1); [ -n "$f" ] || exit 1; printf "%s" "${f#/data/}"' \
+    2>/dev/null) || return 1
+  [ -n "$rel" ] || return 1
+  # Reuse the mirror's own digest-pinned image for the probe Pod. It is the same
+  # image many CI hooks share, so by this late stage of the suite it is usually
+  # already cached on whatever node the probe lands on and no pull is needed; if
+  # the probe does land on a node without it, the pull is bounded by the Pod's
+  # --pod-running-timeout below and a timeout just falls back to the public factory.
+  img=$(kubectl -n kube-system get deploy talos-image-cache \
+    -o jsonpath='{.spec.template.spec.containers[0].image}' 2>/dev/null) || return 1
+  [ -n "$img" ] || return 1
+
+  # --retry rides out the seconds Cilium needs to program the freshly-applied allow
+  # policy; a persistently unreachable ClusterIP still fails fast per attempt
+  # (connect-timeout) and returns a non-206, falling back to the public factory.
+  # The total budget is deliberately generous: this decision is cached once for the
+  # whole suite, so a single slow policy-programming window must not permanently
+  # disable the mirror. --max-time still bounds a hung attempt (a genuinely
+  # unreachable ClusterIP resolves well within it and falls back).
+  curl_cmd="curl -s -o /dev/null -w code=%{http_code} --retry 8 --retry-all-errors --retry-delay 5 --connect-timeout 5 --max-time 45 -r 0-0 ${TALOS_IMAGE_CACHE_SVC_URL}/${rel}"
+  overrides=$(_talos_image_cache_probe_overrides "$img" "$curl_cmd")
+
+  kubectl -n "$TALOS_IMAGE_CACHE_PROBE_NS" delete pod "$TALOS_IMAGE_CACHE_PROBE_POD" \
+    --ignore-not-found >/dev/null 2>&1
+  # No --rm: with --rm the Pod is gone the instant it exits, and --attach can lose
+  # the tail of a fast-completing Pod's stream, leaving $out empty — a false negative
+  # that would fall back to the flaky public factory this mirror exists to avoid.
+  # Keep the Pod so its logs can be re-read if the attach stream was lost, then
+  # delete it explicitly below.
+  out=$(kubectl -n "$TALOS_IMAGE_CACHE_PROBE_NS" run "$TALOS_IMAGE_CACHE_PROBE_POD" \
+    --restart=Never --attach --pod-running-timeout=90s \
+    --image="$img" --labels='cdi.kubevirt.io=importer' \
+    --overrides="$overrides" 2>&1) || true
+  # --attach returns only after the container exits, so its logs are ready now;
+  # fall back to them when the attach stream carried no result marker.
+  if ! printf '%s' "$out" | grep -q 'code='; then
+    out=$(kubectl -n "$TALOS_IMAGE_CACHE_PROBE_NS" logs "$TALOS_IMAGE_CACHE_PROBE_POD" 2>/dev/null || true)
+  fi
+  # Fire-and-forget: the result is already captured in $out, so nothing depends on
+  # the Pod actually being gone. The pre-run delete above must NOT do this — there,
+  # `kubectl run` would race a still-terminating Pod of the same name.
+  kubectl -n "$TALOS_IMAGE_CACHE_PROBE_NS" delete pod "$TALOS_IMAGE_CACHE_PROBE_POD" \
+    --ignore-not-found --wait=false >/dev/null 2>&1
+  _talos_image_cache_probe_succeeded "$out"
 }
 
 # resolve_talos_image_factory_url: print the imageFactoryURL to use, or an empty
@@ -44,21 +153,21 @@ resolve_talos_image_factory_url() {
   local url=""
   if kubectl -n kube-system get deploy talos-image-cache >/dev/null 2>&1; then
     # Available only after the seed initContainer has fully fetched the image, so
-    # this is the signal that the mirror is warm and safe to point tenants at.
+    # this is the signal that the mirror is warm.
     if kubectl -n kube-system rollout status deploy/talos-image-cache --timeout=12m >/dev/null 2>&1; then
-      url="$TALOS_IMAGE_CACHE_SVC_URL"
-      # Available only proves the seed finished and the port is open. Verify the
-      # server actually answers a byte-range request with 206 (not a range-less
-      # 200): range support is what lets CDI skip the slow copy-to-scratch path,
-      # and a future swap to a range-incapable server would otherwise regress it
-      # silently. Probe from inside the serve pod so this does not depend on the
-      # sandbox reaching a ClusterIP. A non-206 still imports correctly (just
-      # without the fast path), so warn loudly rather than discard the mirror and
-      # reintroduce the public factory's egress flakiness.
-      if _talos_image_cache_serves_ranges; then
-        echo "talos-image-cache mirror ready (byte-range verified) — tenant workers import from ${url}" >&2
+      # Available proves only that the seed finished and the server answers on
+      # localhost — NOT that a worker's importer, penned in by the tenant egress
+      # default-deny, can reach the ClusterIP. Install the allow policy (Cilium is
+      # up by now, unlike at seed-deploy time) and verify the whole importer path
+      # end to end. Point tenants at the mirror only when that check passes;
+      # otherwise fall back to the public factory so the mirror can only help,
+      # never point workers at an unreachable ClusterIP and make CI worse.
+      _apply_talos_image_cache_egress_policy || true
+      if _talos_image_cache_reachable_from_tenant; then
+        url="$TALOS_IMAGE_CACHE_SVC_URL"
+        echo "talos-image-cache mirror reachable from tenant namespace (byte-range 206 verified) — tenant workers import from ${url}" >&2
       else
-        echo "WARNING: talos-image-cache is Available but did not answer a range request with 206 — tenant workers still import from ${url}, but CDI will fall back to a full copy" >&2
+        echo "WARNING: talos-image-cache is Available but a tenant-scoped probe could not reach its ClusterIP with a 206 — tenant workers fall back to public factory.talos.dev" >&2
       fi
     else
       echo "WARNING: talos-image-cache mirror not Available in time — tenant workers fall back to public factory.talos.dev" >&2
diff --git a/hack/e2e-install-cozystack.bats b/hack/e2e-install-cozystack.bats
index 627c20c14..9e488536a 100644
--- a/hack/e2e-install-cozystack.bats
+++ b/hack/e2e-install-cozystack.bats
@@ -42,11 +42,12 @@
   # (see cozystack/cozystack#3231). This Deployment fetches the image ONCE with a
   # hard curl retry loop and serves it locally; tenant CRs then point
   # spec.talos.imageFactoryURL at its Service. Deployed here (before the long
-  # install) so the seed download overlaps the install churn — readiness is gated
-  # at point-of-use in hack/e2e-apps/run-kubernetes.sh, which falls back to the
-  # public factory if the mirror never becomes Available, so this can only help.
-  # Best-effort: never fail the suite on the band-aid. Remove once tenant workers
-  # no longer bulk-pull the OS image from the public internet in CI.
+  # install) so the seed download overlaps the install churn — reachability is
+  # gated at point-of-use in hack/e2e-apps/run-kubernetes.sh, which falls back to
+  # the public factory unless a tenant-scoped probe confirms a worker can actually
+  # reach the ClusterIP, so this can only help. Best-effort: never fail the suite
+  # on the band-aid. Remove once tenant workers no longer bulk-pull the OS image
+  # from the public internet in CI.
   local sid ver
   sid=$(yq '.talos.schematicID' packages/apps/kubernetes/values.yaml 2>/dev/null)
   ver=$(yq '.talos.version' packages/apps/kubernetes/values.yaml 2>/dev/null)
@@ -54,7 +55,12 @@
     echo "WARNING: could not read talos.schematicID/version from values.yaml — skipping mirror (fallback to public factory)"
     return 0
   fi
+  # The CiliumClusterwideNetworkPolicy in the manifest is intentionally NOT applied
+  # here: Cilium's CRDs do not exist until Cozystack is installed (below), so
+  # applying it now would error. It is applied at point-of-use by
+  # hack/e2e-apps/talos-image-cache.sh once Cilium is up.
   sed -e "s|__SCHEMATIC_ID__|${sid}|g" -e "s|__TALOS_VERSION__|${ver}|g" hack/e2e-talos-image-cache.yaml \
+    | yq 'select(.kind != "CiliumClusterwideNetworkPolicy")' \
     | kubectl apply -f - || echo "WARNING: failed to apply talos-image-cache (fallback to public factory)"
   if kubectl -n kube-system get deploy talos-image-cache >/dev/null 2>&1; then
     echo "talos-image-cache Deployment created (seeding ${sid}/${ver} in background)"
diff --git a/hack/e2e-talos-image-cache.yaml b/hack/e2e-talos-image-cache.yaml
index 0435a6902..8df38b582 100644
--- a/hack/e2e-talos-image-cache.yaml
+++ b/hack/e2e-talos-image-cache.yaml
@@ -16,8 +16,15 @@
 # per-worker imports), then serves it locally over plain HTTP with range support.
 # Tenant `Kubernetes` CRs point `spec.talos.imageFactoryURL` at this Service
 # (see hack/e2e-apps/run-kubernetes.sh) so every worker imports from the warm
-# local copy instead of the public internet. If this Deployment never becomes
-# Available the e2e harness falls back to the public factory (status quo), so the
+# local copy instead of the public internet. Tenant namespaces run under a
+# default-deny Cilium egress installed by the tenant chart, whose allow rules cover
+# same-namespace endpoints, the tenant tree and the `world` entity (which is why the
+# public factory works) but NOT an arbitrary kube-system Service, so a worker's CDI
+# importer would resolve this Service yet have its TCP connect to the ClusterIP
+# silently dropped. The CiliumClusterwideNetworkPolicy at the end of this file punches a
+# single tightly-scoped hole for that path. The e2e harness only points tenants
+# here after a tenant-scoped probe confirms the importer path actually reaches the
+# ClusterIP with a 206; otherwise it falls back to the public factory, so the
 # mirror can only help, never make CI worse.
 #
 # The __SCHEMATIC_ID__ / __TALOS_VERSION__ placeholders are substituted at apply
@@ -221,3 +228,51 @@ data:
     with http.server.ThreadingHTTPServer(("0.0.0.0", PORT), RangeHandler) as httpd:
         print("serving %s on :%d" % (ROOT, PORT), flush=True)
         httpd.serve_forever()
+---
+# CI-only: allow the tenant CDI importer Pods to reach the mirror.
+#
+# Tenant namespaces run under a default-deny Cilium egress installed by the tenant
+# chart (allow-internal-communication, allow-external-communication and the
+# `<tenant>-egress` CiliumClusterwideNetworkPolicy). Their union allows egress to
+# same-namespace endpoints, the tenant tree and the `world` entity (so the public
+# factory works), but nothing permits egress to an arbitrary kube-system Service —
+# so a worker's CDI importer resolves talos-image-cache.kube-system.svc yet its TCP
+# connect to the Service ClusterIP is silently dropped (dial ... i/o timeout), the
+# disk import never starts, and the tenant node never joins. Cilium unions allow
+# rules across policies, so this adds one tightly-scoped allow: importer Pods in
+# tenant-test may egress to the mirror Pod in kube-system.
+#
+# Scoped to tenant-test (which the tenant chart already puts under default-deny
+# egress) so it only ever ADDS an allow — it never flips another namespace from
+# default-allow to default-deny. The mirror Pod is selected by no ingress policy
+# (default-allow ingress), so no ingress rule is needed; replies return on the
+# established conntrack entry.
+#
+# Applied at point-of-use by hack/e2e-apps/talos-image-cache.sh, not with the rest
+# of this manifest in hack/e2e-install-cozystack.bats: that runs before Cozystack
+# (and thus before Cilium's CRDs) is installed, so this kind does not yet exist
+# when the manifest is first applied.
+apiVersion: cilium.io/v2
+kind: CiliumClusterwideNetworkPolicy
+metadata:
+  name: e2e-talos-image-cache-importer-egress
+  labels:
+    app.kubernetes.io/name: talos-image-cache
+spec:
+  endpointSelector:
+    matchLabels:
+      "k8s:cdi.kubevirt.io": importer
+      "k8s:io.kubernetes.pod.namespace": tenant-test
+  # No toPorts: the allow is already scoped to a single kube-system pod identity
+  # (CI-only, torn down with the sandbox), so an L4 port restriction adds no
+  # meaningful blast-radius reduction — only risk. With kube-proxy-less Cilium the
+  # importer's connect to the Service ClusterIP is socket-LB translated to the
+  # backend pod port BEFORE egress policy is evaluated, so a port literal here
+  # would have to track the container port (8080), not the Service port (80); a
+  # wrong guess silently drops the traffic and reintroduces the exact public-factory
+  # fallback this policy exists to prevent.
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": kube-system
+            "k8s:app.kubernetes.io/name": talos-image-cache
diff --git a/hack/talos-image-cache_test.bats b/hack/talos-image-cache_test.bats
new file mode 100644
index 000000000..9a3225d35
--- /dev/null
+++ b/hack/talos-image-cache_test.bats
@@ -0,0 +1,144 @@
+#!/usr/bin/env bats
+# -----------------------------------------------------------------------------
+# Unit tests for the in-sandbox Talos image cache manifest split and the
+# importer-reachability contract it relies on.
+#
+# hack/e2e-talos-image-cache.yaml bundles four documents but they are applied in
+# two phases: hack/e2e-install-cozystack.bats applies everything EXCEPT the
+# CiliumClusterwideNetworkPolicy (its CRD does not exist before Cozystack is
+# installed), and hack/e2e-apps/talos-image-cache.sh applies ONLY that policy
+# later, once Cilium is up. If a future document is added to the manifest and
+# silently dropped from the pre-Cilium apply, or the Cilium document leaks into
+# it (and errors on the missing CRD), the mirror breaks and e2e falls back to
+# the flaky public factory. These tests pin that split.
+#
+# They also pin the load-bearing reachability invariant: the throwaway probe pod
+# is a faithful proxy for a real CDI importer only if it carries the exact label
+# and namespace the egress policy selects. If either side drifts, the probe can
+# pass while real importers stay blocked (a false positive that makes CI worse).
+#
+# Finally they cover the two pure fragments of the decision itself, sourced from
+# hack/e2e-apps/talos-image-cache.sh (sourcing it has no cluster side effects — it
+# only sets parameter defaults and defines functions): the strict 206 gate that
+# decides whether tenants are pointed at the mirror at all, and the --overrides JSON
+# builder whose silent breakage would drop every probe to the public factory. The
+# kubectl orchestration around them needs a live cluster and is covered by the e2e
+# run, matching how the other hack/ helpers are tested.
+#
+# cozytest.sh's awk parser recognizes only @test blocks and a bare `}` on its
+# own line; there is no bats `run` or `$status`, and setup()/teardown() are not
+# honored. Each test runs under `set -eu -x`; assertions are direct shell tests
+# that exit non-zero on failure. mikefarah yq prints `---` between matched
+# documents, so document streams are compared with those separators stripped.
+#
+# Run with: hack/cozytest.sh hack/talos-image-cache_test.bats
+# -----------------------------------------------------------------------------
+
+@test "manifest documents partition into pre-Cilium apply plus the Cilium policy" {
+    manifest=hack/e2e-talos-image-cache.yaml
+    total=$(yq '.kind' "$manifest" | grep -vc '^---$')
+    excluded=$(yq 'select(.kind != "CiliumClusterwideNetworkPolicy") | .kind' "$manifest" | grep -vc '^---$')
+    selected=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .kind' "$manifest" | grep -vc '^---$')
+    [ "$total" -eq 4 ]
+    [ "$excluded" -eq 3 ]
+    [ "$selected" -eq 1 ]
+    [ $((excluded + selected)) -eq "$total" ]
+}
+
+@test "pre-Cilium apply keeps Service, Deployment, ConfigMap and drops the Cilium policy" {
+    manifest=hack/e2e-talos-image-cache.yaml
+    kinds=$(yq 'select(.kind != "CiliumClusterwideNetworkPolicy") | .kind' "$manifest" | grep -v '^---$')
+    for want in Service Deployment ConfigMap; do
+        printf '%s\n' "$kinds" | grep -qx "$want" || { echo "pre-Cilium apply is missing $want" >&2; exit 1; }
+    done
+    if printf '%s\n' "$kinds" | grep -qx CiliumClusterwideNetworkPolicy; then
+        echo "CiliumClusterwideNetworkPolicy leaked into the pre-Cilium apply" >&2
+        exit 1
+    fi
+}
+
+@test "point-of-use apply selects exactly the Cilium policy" {
+    manifest=hack/e2e-talos-image-cache.yaml
+    kinds=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .kind' "$manifest" | grep -v '^---$')
+    [ "$kinds" = "CiliumClusterwideNetworkPolicy" ]
+}
+
+@test "egress policy selects the importer label and namespace the probe pod uses" {
+    manifest=hack/e2e-talos-image-cache.yaml
+    helper=hack/e2e-apps/talos-image-cache.sh
+    label=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .spec.endpointSelector.matchLabels["k8s:cdi.kubevirt.io"]' "$manifest")
+    ns=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .spec.endpointSelector.matchLabels["k8s:io.kubernetes.pod.namespace"]' "$manifest")
+    [ "$label" = "importer" ]
+    [ "$ns" = "tenant-test" ]
+    # The probe pod must carry the same label the policy selects, else it faces a
+    # different egress than a real importer and the guarantee breaks.
+    grep -q "cdi.kubevirt.io=importer" "$helper" || { echo "probe pod label drifted from the egress selector" >&2; exit 1; }
+    grep -q "TALOS_IMAGE_CACHE_PROBE_NS:-tenant-test" "$helper" || { echo "probe namespace drifted from the egress selector" >&2; exit 1; }
+}
+
+@test "egress rule targets the mirror pod's own identity" {
+    manifest=hack/e2e-talos-image-cache.yaml
+    target=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .spec.egress[0].toEndpoints[0].matchLabels["k8s:app.kubernetes.io/name"]' "$manifest")
+    target_ns=$(yq 'select(.kind == "CiliumClusterwideNetworkPolicy") | .spec.egress[0].toEndpoints[0].matchLabels["k8s:io.kubernetes.pod.namespace"]' "$manifest")
+    mirror=$(yq 'select(.kind == "Deployment") | .spec.template.metadata.labels["app.kubernetes.io/name"]' "$manifest")
+    mirror_ns=$(yq 'select(.kind == "Deployment") | .metadata.namespace' "$manifest")
+    [ "$target" = "talos-image-cache" ]
+    [ "$target_ns" = "kube-system" ]
+    # The allow's destination must equal the mirror Deployment's own pod label and
+    # namespace, otherwise the hole is punched toward nothing: moving the mirror
+    # would leave these tests green while importers could no longer reach it.
+    [ "$target" = "$mirror" ]
+    [ "$target_ns" = "$mirror_ns" ]
+}
+
+@test "strict 206 gate selects the mirror on a byte-range success" {
+    . hack/e2e-apps/talos-image-cache.sh
+    if ! _talos_image_cache_probe_succeeded "code=206"; then
+        echo "expected a 206 probe result to select the mirror" >&2
+        exit 1
+    fi
+}
+
+@test "strict 206 gate finds the marker inside surrounding probe output" {
+    . hack/e2e-apps/talos-image-cache.sh
+    out=$(printf 'pod/talos-image-cache-probe created\ncode=206\npod "talos-image-cache-probe" deleted\n')
+    if ! _talos_image_cache_probe_succeeded "$out"; then
+        echo "expected 206 to be found in multi-line attach output" >&2
+        exit 1
+    fi
+}
+
+@test "strict 206 gate falls back on range-less, unreachable, empty and unrelated output" {
+    . hack/e2e-apps/talos-image-cache.sh
+    # A plain 200 means the mirror answered without range support; 000 means curl
+    # never connected. Both fall back to the public factory, as does output that
+    # carried no result marker at all (an attach stream lost with no logs to recover).
+    for bad in "code=200" "code=000" "code=404" "" "error: pods not found"; do
+        if _talos_image_cache_probe_succeeded "$bad"; then
+            echo "expected non-206 output to fall back to the public factory: [$bad]" >&2
+            exit 1
+        fi
+    done
+}
+
+@test "probe overrides builder emits valid JSON with the image, command and restricted PSA" {
+    . hack/e2e-apps/talos-image-cache.sh
+    img=example.invalid/img:tag
+    cmd='curl -s -o /dev/null -w code=%{http_code} -r 0-0 http://talos-image-cache.kube-system.svc/x.raw.xz'
+    json=$(_talos_image_cache_probe_overrides "$img" "$cmd")
+    # Invalid JSON would make every probe fail and silently drop CI back to the
+    # public factory, so parse it rather than pattern-match it.
+    printf '%s' "$json" | yq -p=json -o=json '.' >/dev/null
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].name')" = "c" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].image')" = "$img" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].command[2]')" = "$cmd" ]
+    # Every field the tenant namespace's restricted Pod Security Admission requires:
+    # drop any one of them and the probe Pod is rejected at admission, the probe
+    # fails, and CI silently falls back to the public factory.
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.securityContext.runAsNonRoot')" = "true" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.securityContext.runAsUser')" = "1000" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.securityContext.seccompProfile.type')" = "RuntimeDefault" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].securityContext.allowPrivilegeEscalation')" = "false" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].securityContext.capabilities.drop | length')" = "1" ]
+    [ "$(printf '%s' "$json" | yq -p=json '.spec.containers[0].securityContext.capabilities.drop[0]')" = "ALL" ]
+}