diff --git a/k8s/bases/infrastructure/controllers/flux-operator/helm-release.yaml b/k8s/bases/infrastructure/controllers/flux-operator/helm-release.yaml
index 41deabeff..f87148a58 100644
--- a/k8s/bases/infrastructure/controllers/flux-operator/helm-release.yaml
+++ b/k8s/bases/infrastructure/controllers/flux-operator/helm-release.yaml
@@ -73,12 +73,23 @@ spec:
   # Neither carries the runAsGroup-without-runAsUser coupling documented above, so
   # neither can stop the sandbox building.
   #
-  # 🔴 The pod-level op adds a LEAF into the chart's pod securityContext, which renders
-  # as `{}` (read from the live Deployment 2026-09-04). A leaf add is deliberate: adding
-  # the whole object would REPLACE it and silently discard any field the chart starts
-  # setting later, which is exactly what the +(key) anchors elsewhere exist to prevent.
-  # The cost is that a chart that stopped rendering the object would fail this patch —
-  # loudly, on a Renovate chart bump with CI in front of it, never silently.
+  # 🔴 The two fields need DIFFERENT patch kinds, because the chart renders their
+  # parents differently. Verified by rendering flux-operator 0.50.0 with this
+  # HelmRelease's own values 2026-09-04:
+  # - containers[0].securityContext IS rendered, so a JSON-6902 leaf `add` applies.
+  # - spec.template.spec.securityContext is NOT rendered AT ALL, so a JSON-6902 leaf
+  #   `add` under it fails `doc is missing path` — and because a JSON patch is
+  #   all-or-nothing, that one op takes the container ops down with it.
+  #
+  # Do NOT read this off the live Deployment: the API server serialises the absent
+  # pod-level object as `{}`, so the cluster shows a parent that the rendered chart
+  # does not have. The rendered manifest is the surface a post-renderer patches.
+  #
+  # The pod-level field is therefore a strategic-merge patch, which creates the
+  # missing parent AND merges into a chart-set one — so if a future chart starts
+  # rendering spec.template.spec.securityContext, its fields survive. A JSON-6902
+  # `add` of the whole object would have replaced them silently; that is why this is
+  # a merge rather than an object add.
   postRenderers:
     - kustomize:
         patches:
@@ -99,6 +110,16 @@ spec:
                 path: /spec/template/spec/containers/0/securityContext/seLinuxOptions
                 value:
                   level: s0
-              - op: add
-                path: /spec/template/spec/securityContext/fsGroupChangePolicy
-                value: OnRootMismatch
+          - target:
+              kind: Deployment
+              name: flux-operator
+            patch: |
+              apiVersion: apps/v1
+              kind: Deployment
+              metadata:
+                name: flux-operator
+              spec:
+                template:
+                  spec:
+                    securityContext:
+                      fsGroupChangePolicy: OnRootMismatch
diff --git a/scripts/validate-eks-ci-role-policy/main.go b/scripts/validate-eks-ci-role-policy/main.go
index c8bdfe579..00f692319 100644
--- a/scripts/validate-eks-ci-role-policy/main.go
+++ b/scripts/validate-eks-ci-role-policy/main.go
@@ -1684,10 +1684,40 @@ const (
 // rendered with the SAME local renderer; the digest is not, so it comes from the
 // pinned one.
 //
+// Re-approved 2026-09-04 for #3580, which repairs the change the paragraph above
+// approved. That change could not apply at all: the pod-level op targets
+// /spec/template/spec/securityContext/fsGroupChangePolicy, but the flux-operator
+// chart renders no pod-level securityContext, so helm-controller failed the whole
+// post-render with `doc is missing path` and neither field reached the cluster.
+// #3580 moves that one field to a strategic-merge patch, which creates the missing
+// parent; the container ops are unchanged.
+//
+// Conservation re-proven the same way, with ONE renderer on both sides: all five
+// overlays rendered from main d7613b03 and from this branch, compared over the full
+// apiVersion|kind|namespace|name identity in BOTH directions. 552 documents each
+// side — ZERO added, ZERO removed. Per-root counts (126/226/188/8/4) were
+// corroborated against raw document-separator counts, because `yq eval-all`
+// evaluates once across a whole stream and silently reports 1. A negative control
+// that drops a single document reports removed=1 and names it, so the comparison is
+// not vacuous.
+//
+// Four of the five roots render byte-identical. The ENTIRE delta is confined to the
+// flux-operator HelmRelease in the controllers root: 3 lines removed, 13 added, all
+// inside its postRenderers block. No rule, verb, resource, group, subject,
+// ServiceAccount, role reference or policy moved. The digest moves for the same
+// reason as last time — that HelmRelease is itself a selected authorization-capable
+// document — not because privilege changed.
+//
+// The value below is again CI's own computed digest (run 33866934697, head
+// 8c9d1d11), for the same reason: this validator pins kubectl v1.36.2 and refuses
+// any other, and the preparing host has v1.36.1. That run reported exactly one
+// error class — this fingerprint — so nothing else in the authorization surface
+// objected.
+//
 // The previous approved aggregate digest was:
 //
-//	a16fbd577f8155e64aacc02cb624f26d9628d75658ce804215d9d69a3884bf92
-const expectedRenderedSurfaceSHA = "ab05bc2c95924e372c038f878872aa94e3bd81380daa96a283bd7f74e2132a69"
+//	ab05bc2c95924e372c038f878872aa94e3bd81380daa96a283bd7f74e2132a69
+const expectedRenderedSurfaceSHA = "a5dac56d1ee989648670e2bd265b56e574512b37e2e19b8caf0fe4738816d1a4"
 
 // authorizationOverlayPaths lists every independently reconciled production
 // layer where an object can grant privileges to the aws/aws service account.