diff --git a/docs/sync-and-op-log/vector-clocks.md b/docs/sync-and-op-log/vector-clocks.md
index bab29beee7..217cc977c9 100644
--- a/docs/sync-and-op-log/vector-clocks.md
+++ b/docs/sync-and-op-log/vector-clocks.md
@@ -433,3 +433,46 @@ via **Dotted Version Vectors** (bound to server vnodes, not devices),
 **bounded reclaimable client IDs** (needs a registration/retirement protocol),
 or **periodic stable-cut GC** (needs all-to-all clock reporting). None apply to
 the current dumb-relay model.
+
+### Future option: staleness-informed eviction (issue #9105 — works in the dumb-relay model)
+
+Pruning today evicts the **lowest-counter** entries, but a low counter
+correlates with _importance_ (a fresh import author has counter 1), not with
+_deadness_ — the heuristic behind the #9089/#9096 preserve-set bugs. Issue
+#9105 tracks the root cause: client IDs are minted per install/profile and
+retired almost never, so clocks only grow toward MAX. The decision on #9105
+was to **park** the fix — post #9089/#9102 the worst case is the benign extra
+round-trip of §5 — and record the agreed direction here.
+
+If pruning stops being rare in practice, evict the **stalest** entries instead
+of the lowest-counter ones. Unlike the coordinator options above, this fits
+the dumb-relay model with no wire-format change:
+
+- **Server:** the `sync_devices` table already stores `lastSeenAt` per
+  `(userId, clientId)`, updated on every upload — and uploads are the only
+  path that creates clock entries. A daily job already GCs rows unseen for
+  `retentionMs` (45 days), so absence from the registry reads as "stalest".
+- **Client (all providers):** keep a small durable `clientId → last-merged-op
+time` map, updated where remote clocks are merged (`mergeRemoteOpClocks`) —
+  every merged op carries its author's ID. Needs no server support, so it
+  covers WebDAV / LocalFile / Dropbox too.
+
+The safety profile is identical to today's pruning (entries are dropped either
+way; a dropped ID that returns costs at most the extra round-trip of §5), but
+victim selection is strictly better: a recently-seen ID — e.g. a fresh import
+author — survives by definition, making the preserve-set invariant of
+#9089/#9096 _emergent_ instead of hand-maintained at each prune site (the
+explicit preserve sets stay as belt-and-braces). Staleness knowledge differs
+per node, so nodes may evict different victims; that adds clock asymmetry but
+no new failure class — comparison treats missing keys as zero, and clients
+already prune with differing preserve sets.
+
+The supported GC today is a **full-state import**: the clock reset keeps only
+`{import author, self}` (§7), and the once-per-session pruning snack points
+users at it (sync all devices first — imports intentionally drop concurrent
+ops, see `SyncImportFilterService`).
+
+**Revisit trigger:** client pruning WARN-logs `prunedIds` / `survivingIds`
+into the exportable log history. If prune warnings appear in real bug
+reports — especially ones evicting _live_ IDs — promote this from parked to
+scheduled.
diff --git a/src/app/imex/sync/sync.effects.ts b/src/app/imex/sync/sync.effects.ts
index 80326e09ed..e18d640358 100644
--- a/src/app/imex/sync/sync.effects.ts
+++ b/src/app/imex/sync/sync.effects.ts
@@ -119,15 +119,15 @@ export class SyncEffects {
           this._snackService.open({
             msg: T.F.SYNC.S.VECTOR_CLOCK_LIMIT_REACHED,
             // CUSTOM (not WARNING): this is a benign, self-healing cleanup, not
-            // an alert the user must act on. Neutral icon, auto-dismiss — but a
-            // touch longer than the 3s default so the sentence is readable.
+            // an alert the user must act on. Neutral icon, auto-dismiss — long
+            // enough to read the optional clear-for-good remedy (#9105).
             type: 'CUSTOM',
             ico: 'sync',
             translateParams: {
               originalSize,
               maxSize,
             },
-            config: { duration: 5000 },
+            config: { duration: 8000 },
           });
         }),
       ),
diff --git a/src/assets/i18n/en.json b/src/assets/i18n/en.json
index 4df195b9e7..711436ac77 100644
--- a/src/assets/i18n/en.json
+++ b/src/assets/i18n/en.json
@@ -1668,7 +1668,7 @@
         "UNKNOWN_ERROR": "Unknown Sync Error: {{err}}",
         "UPLOAD_ERROR": "Unknown Upload Error (Settings correct?): {{err}}",
         "UPLOAD_OPS_REJECTED": "{{count}} operation(s) were permanently rejected by the server and won't be synced.",
-        "VECTOR_CLOCK_LIMIT_REACHED": "Sync tidied up old device entries ({{originalSize}} → {{maxSize}}). This is normal — no action needed.",
+        "VECTOR_CLOCK_LIMIT_REACHED": "Sync tidied up old device entries ({{originalSize}} → {{maxSize}}). This is normal. To clear them for good: sync all devices, then export and re-import a backup (Settings → Import/Export).",
         "VERSION_TOO_OLD": "Your app version is too old for the synced data. Please update!",
         "VERSION_UNSUPPORTED": "Cannot sync: some synced data is from an app version that is no longer supported. Sync is paused before it; updating the app on your other devices may help.",
         "WEB_CRYPTO_NOT_AVAILABLE": "Encryption is not available on this device. On Android, encryption requires a secure context (HTTPS) which is not available. Please disable encryption in sync settings or sync from a desktop web browser."