diff --git a/CHANGELOG.md b/CHANGELOG.md
index a820ed4..7497ff2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -32,6 +32,16 @@ uses semantic versioning (`MAJOR.MINOR.PATCH`).
 
 ### Fixed
 
+- **Watch: a queued offline switch can no longer be created twice.** The watch
+  app, its data-layer service and its tile trampoline each drained the offline
+  switch queue with their own lock, so two of them could submit the same queued
+  switch at once. They now share one lock across the whole app.
+
+- **Phone: syncing offline actions no longer risks duplicating a front.** A new
+  sync could previously cancel one already in progress at the moment it was
+  creating a front, which could create it twice; a running sync is now left to
+  finish.
+
 - **Signing into another account no longer leaks the previous account's data.**
   Signing in or out now clears the local response cache and the offline action
   queue, so a new account can't briefly see the prior account's members, fronts
diff --git a/sheaf/app/src/main/java/systems/lupine/sheaf/data/sync/SyncWorker.kt b/sheaf/app/src/main/java/systems/lupine/sheaf/data/sync/SyncWorker.kt
index 043c66d..47a8813 100644
--- a/sheaf/app/src/main/java/systems/lupine/sheaf/data/sync/SyncWorker.kt
+++ b/sheaf/app/src/main/java/systems/lupine/sheaf/data/sync/SyncWorker.kt
@@ -121,9 +121,19 @@ class SyncWorker @AssistedInject constructor(
                 .setConstraints(Constraints(requiredNetworkType = NetworkType.CONNECTED))
                 .setBackoffCriteria(BackoffPolicy.EXPONENTIAL, 15, java.util.concurrent.TimeUnit.SECONDS)
                 .build()
+            // KEEP, not REPLACE: a drain that is already running must not be
+            // cancelled by a fresh enqueue. createFront is posted before its row
+            // is deleted (see doWork), so a REPLACE cancellation mid-drain is
+            // exactly the window that can duplicate a front. With KEEP, an
+            // enqueue while a run is in flight is dropped rather than cancelling
+            // it; any rows added after that run snapshotted the queue are picked
+            // up by the next schedule() (each offline action, plus reconnect,
+            // calls this). Trading a little latency on the tail for not
+            // double-creating fronts. True exactly-once still needs the backend
+            // Idempotency-Key noted in doWork.
             WorkManager.getInstance(context).enqueueUniqueWork(
                 WORK_NAME,
-                ExistingWorkPolicy.REPLACE,
+                ExistingWorkPolicy.KEEP,
                 request,
             )
         }
diff --git a/sheaf/wear/src/main/java/systems/lupine/sheaf/wear/data/WearStore.kt b/sheaf/wear/src/main/java/systems/lupine/sheaf/wear/data/WearStore.kt
index 1150ecb..c1bb338 100644
--- a/sheaf/wear/src/main/java/systems/lupine/sheaf/wear/data/WearStore.kt
+++ b/sheaf/wear/src/main/java/systems/lupine/sheaf/wear/data/WearStore.kt
@@ -33,8 +33,9 @@ class WearStore(
     // Serialises the offline-switch queue drain. loadAll()/refreshNow() run from
     // many triggers (onResume, nav, manual refresh, post-switch, phone nudge);
     // without this, two concurrent drains could both submit the same queued row
-    // before either removed it, double-creating a front.
-    private val queueDrainMutex = Mutex()
+    // before either removed it, double-creating a front. The mutex lives in the
+    // companion object (see below) so it is shared across every WearStore in the
+    // process.
 
     val frontingMembers: List<WearMember>
         get() {
@@ -278,6 +279,16 @@ class WearStore(
         // or the complications aren't currently in use, no harm done.
         systems.lupine.sheaf.wear.complications.requestAllComplicationUpdates(context)
     }
+
+    companion object {
+        // Process-wide, not per-instance: the activity, the Data Layer service
+        // and the quick-switch trampoline each build their own WearStore, but
+        // they drain one shared SharedPreferences queue. A per-instance mutex
+        // let two of them snapshot and re-submit the same row, double-creating a
+        // front. Mirrors WearApiClient.refreshMutex, which was hoisted for the
+        // same reason.
+        private val queueDrainMutex = Mutex()
+    }
 }
 
 // ── Shared tile_data writers ────────────────────────────────────────────────