fix: prevent stale snapshots resurrecting orphaned dashboards

CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## v1.8.22-alpha (2026-09-30)
 
+- Network diagnostic failures no longer stop all apps or rebuild shared container networking.
+- Prevented orphaned companion dashboards from repeatedly reinstalling themselves after their backend app was removed.
+- Fixed companion dashboard builds still referencing a retired image registry.
+
 - Fixed Angor Indexer health checks choosing IPv6 localhost for an IPv4 listener and unnecessarily restarting the working service.
 
 - Prevented false app restarts by probing each published port at its actual bind address; Nginx Proxy Manager now checks its internal admin API.

core/archipelago/src/container/companion.rs

@@ -103,6 +103,15 @@ pub fn companions_for(package_id: &str) -> &'static [CompanionSpec] {
     }
 }
 
+/// Missing companion UIs are provisioned here, never by snapshot recovery.
+/// A stale running-container snapshot must not resurrect an orphaned UI.
+pub fn is_companion_app(app_id: &str) -> bool {
+    ALL_COMPANIONS
+        .iter()
+        .flat_map(|specs| specs.iter())
+        .any(|spec| spec.image_base == app_id)
+}
+
 /// Every companion this build knows how to provision. Kept beside
 /// `companions_for` — a new companion must be added to both, or the reaper
 /// will not recognise it as one of ours and will leave it running forever.

core/archipelago/src/container/prod_orchestrator.rs

@@ -2071,6 +2071,10 @@ impl ProdContainerOrchestrator {
                 Ok(ReconcileAction::Left(reason))
                     if mode == ReconcileMode::ExistingOnly
                         && reason == "absent"
+                        // companion.rs owns missing UI provisioning/removal.
+                        // Never resurrect an orphan from a stale snapshot.
+                        // Existing UIs still pass through security config repair.
+                        && !super::companion::is_companion_app(&app_id)
                         && (was_running.contains(&compute_container_name(&lm.manifest))
                             // The durable answer, and the one that does not
                             // erode. `was_running` only records what was
@@ -7225,6 +7229,52 @@ app:
         assert!(!calls.iter().any(|c| c.starts_with("start_container:")));
     }
 
+    #[tokio::test]
+    async fn reconcile_existing_does_not_resurrect_orphaned_companions() {
+        let rt = Arc::new(MockRuntime::default());
+        let mut orch = orch_with(rt.clone()).await;
+        orch.set_disk_gb_for_test(500);
+        let companions = [
+            "bitcoin-ui",
+            "electrs-ui",
+            "lnd-ui",
+            "fedimint-ui",
+            "cuprate-ui",
+        ];
+        let mut names = Vec::new();
+        for id in companions {
+            let manifest = pull_manifest(id, "localhost/companion:local");
+            names.push(compute_container_name(&manifest));
+            orch.insert_manifest_for_test(manifest, PathBuf::from("/tmp/companion"))
+                .await;
+        }
+        let refs: Vec<&str> = names.iter().map(String::as_str).collect();
+        crate::crash_recovery::save_container_snapshot_for_test(&orch.data_dir, &refs).await;
+        // Repeated passes must leave lifecycle ownership with companion.rs.
+        for _ in 0..3 {
+            let report = orch.reconcile_existing().await;
+            assert_eq!(report.actions.len(), companions.len());
+            assert!(report
+                .actions
+                .iter()
+                .all(|(_, action)| *action == ReconcileAction::Left("absent".into())));
+            assert!(report.failures.is_empty());
+        }
+        let calls = rt.calls();
+        for operation in [
+            "pull_image:",
+            "create_container:",
+            "start_container:",
+            "stop_container:",
+            "remove_container:",
+        ] {
+            assert!(
+                !calls.iter().any(|call| call.starts_with(operation)),
+                "{calls:?}"
+            );
+        }
+    }
+
     #[tokio::test]
     async fn reconcile_existing_self_heals_missing_optional_installed_app() {
         // A non-baseline app (gitea) self-heals ONLY with installation

docs/next-release-20260930.md

@@ -334,3 +334,54 @@ repository branch and Compose content from its own network namespace.
 Version preparation is 1.8.22-alpha. No new release tag or fleet-visible update
 manifest is published by the version commit. Optimized candidate deployment,
 artifact inspection, ISO smoke/boot checks and offline signatures follow.
+
+### Release blocker discovered during candidate observation: scheduled doctor
+
+The initial `02b840f2` 1.8.22 candidate is rejected for release. On the X250,
+2026-09-30 21:10–21:11 UTC, the scheduled `archipelago-doctor.service` explicitly
+ran `podman stop --all --time 30`, killed rootless network helpers and ran
+`podman system migrate` after a two-attempt external network probe failed.
+The journal attributes the stop to that unit, not the app health monitor or a
+host reboot. All apps restarted, including Bitcoin, LND and the production site.
+The earlier unchanged-container acceptance applies only to immediate deployment;
+the later observation failed and must not be represented as a stability pass.
+No persistent-data loss has been established. Keep this distinct from the closed
+Framework incident; do not wipe or recreate any wallet as a recovery action.
+
+Containment: stopped doctor timers on both test boxes, installed a safe diagnostic
+script into both the executable and runtime payload, and rejected/stopped the
+old ISO build. Network failure now produces a warning without stopping apps,
+killing network processes, migrating Podman or deleting network state. Repeated
+failures remain warnings, never a successful repair/check. Regression cases cover
+healthy, absent network, non-root invocation, host failure, transient recovery,
+repeated endpoint failure and namespace access failure, with mutation tripwires.
+Live scheduled-cycle observation and final rebuilt-artifact acceptance are pending.
+
+Recovery also exposed retired `git.tx1138.com` nginx base references in six
+companion UI Dockerfiles. They now use the existing primary registry at the same
+pinned version. All six images built successfully against that registry; payload
+validation rejects the retired host before OTA/ISO packaging.
+
+The post-recovery X250 check passes: Bitcoin authenticated RPC responds and IBD
+advances; NPM/Gitea/Portainer APIs respond; Portainer's real namespace fetches
+`demo-portainer` at `3ae171d6b0c728665a860520fe393c0abb772798` and its Compose
+file; Portainer's original persistent mounts match the earlier backup evidence;
+LND wallet/channel databases remain present on their persistent mount. No new
+pre-incident cryptographic wallet-identity baseline was available, so these checks
+must not be described as an exact identity/balance comparison.
+
+The dev all-container observation also caught a separate Cuprate UI orphan loop:
+`companion.rs` removed it because Cuprate was not installed, while generic desired-
+state recovery resurrected it from an old running snapshot, using a unit without
+nginx's required capabilities. Generic desired-state recovery now excludes missing companions
+owned by `companion.rs`; existing companion provisioning/reaping remains the
+single owner. Running UIs still receive the existing security configuration repairs. Regression runs repeated reconciliation against stale companion
+snapshots and checks that no image/container lifecycle operations occur.
+
+Safe-doctor live acceptance: the X250 completed a 12-minute observation with all
+running container IDs, start times and data mounts unchanged. Its journal records
+successful doctor runs at 21:22:06, 21:27:51 and 21:33:10 UTC. Both doctor timers
+are restored with the safe script. Dev's native Bitcoin/LND stayed running;
+all-container dev acceptance remains pending the companion-loop backend fix.
+Final-source UI suite: 1,133 passed. Heavy backend compilation is serialized with
+remaining build steps to reduce memory/IO pressure on the syncing dev node.

neode-ui/src/views/settings/AccountInfoSection.vue

@@ -369,6 +369,9 @@ init()
                   <span class="text-xs text-white/40">September 30, 2026</span>
                 </div>
                 <div class="space-y-3 text-sm text-white/80 pl-3 border-l border-white/10">
+                  <p>Network diagnostic failures no longer stop all apps or rebuild shared container networking.</p>
+                  <p>Prevented orphaned companion dashboards from repeatedly reinstalling themselves after their backend app was removed.</p>
+                  <p>Fixed companion dashboard builds still referencing a retired image registry.</p>
                   <p>Fixed Angor Indexer health checks choosing IPv6 localhost for an IPv4 listener and unnecessarily restarting the working service.</p>
                   <p>Prevented false app restarts by probing each published port at its actual bind address; Nginx Proxy Manager now checks its internal admin API.</p>
                   <p>Added a backed-up migration for the recognized legacy Nginx Proxy Manager tunnel/LND port conflict in both OTA and ISO startup paths.</p>