diff --git a/docs/runbooks/restore-the-stack.md b/docs/runbooks/restore-the-stack.md
index ecb2213..b4e1f4f 100644
--- a/docs/runbooks/restore-the-stack.md
+++ b/docs/runbooks/restore-the-stack.md
@@ -287,13 +287,16 @@ started on the result; §4 was run against it, including the negative assertions
 
 **What it found, which per-volume testing had not.**
 
-- **Grafana will not start without internet access.** `GF_INSTALL_PLUGINS` makes
-  the background installer contact `grafana.com` on every start, and a failure
-  there is fatal — Grafana crash-loops, even though both plugins are already
-  present in the restored volume. On a host that has lost its uplink, which is a
-  perfectly ordinary disaster, the restore succeeds and Grafana still will not
-  come up. Clearing that variable is the workaround; the plugins in the volume
-  are used regardless.
+- **Grafana would not start without internet access.** `GF_INSTALL_PLUGINS`
+  made the background installer contact `grafana.com` on every start, and a
+  failure there was fatal — Grafana crash-looped, even though both plugins were
+  already present in the restored volume. On a host that has lost its uplink,
+  which is a perfectly ordinary disaster, the restore succeeded and Grafana
+  still would not come up. Neither plugin was used by any dashboard, and one of
+  them was an Angular plugin this Grafana refuses to load anyway, so the
+  declaration was removed rather than repaired. Grafana's own bundled apps ship
+  inside the image and need no network, which is why an offline start works now
+  — verified on both fresh and restored volumes.
 - **Alloy replays, and now there is a number for it.** Restoring `alloy-data`
   put the log positions back to their offsets at the stamp, and Alloy re-read
   from there and re-shipped the lines with their original timestamps — about
diff --git a/stacks/observability/compose.yaml b/stacks/observability/compose.yaml
index 83bd7f3..d966e2b 100644
--- a/stacks/observability/compose.yaml
+++ b/stacks/observability/compose.yaml
@@ -180,7 +180,28 @@ services:
       GF_AUTH_ANONYMOUS_ENABLED: "false"
       GF_ANALYTICS_REPORTING_ENABLED: "false"
       GF_ANALYTICS_CHECK_FOR_UPDATES: "false"
-      GF_INSTALL_PLUGINS: grafana-clock-panel,grafana-piechart-panel
+      # No plugins are installed here, deliberately. This was
+      # GF_INSTALL_PLUGINS: grafana-clock-panel,grafana-piechart-panel and both
+      # were dead weight — no dashboard under grafana/dashboards/ uses either
+      # panel type, and grafana-piechart-panel is an Angular plugin, which this
+      # Grafana refuses to initialise: it was downloaded and then rejected on
+      # every single start.
+      #
+      # It was also an availability bug, which is the reason it is gone rather
+      # than merely tidied. The background installer contacts grafana.com on
+      # every start, and a failure there is FATAL to the process — Grafana
+      # crash-looped on a host with no uplink, even though both plugins were
+      # already present in grafana-data. That turns an ordinary outage into a
+      # failed disaster recovery, and it was found by doing one: see the last
+      # section of docs/runbooks/restore-the-stack.md.
+      #
+      # Grafana still preinstalls its own bundled apps (pyroscope, exploretraces,
+      # metricsdrilldown, lokiexplore, elasticsearch). Those ship inside the
+      # image and need no network, which is what makes an offline start work.
+      #
+      # If a plugin is ever genuinely needed, add GF_PLUGINS_PREINSTALL —
+      # GF_INSTALL_PLUGINS is deprecated — and re-test the offline start before
+      # committing it.
       GF_PATHS_PROVISIONING: /etc/grafana/provisioning
       PROMETHEUS_URL: http://prometheus:9090
       LOKI_URL: http://loki:3100