diff --git a/crates/flux-communication/src/queue/mod.rs b/crates/flux-communication/src/queue/mod.rs
index 21f3ef9..de0ecb0 100644
--- a/crates/flux-communication/src/queue/mod.rs
+++ b/crates/flux-communication/src/queue/mod.rs
@@ -575,6 +575,27 @@ impl<T: Copy> Queue<T> {
     fn group_cursor(&self, key: &str) -> *const AtomicUsize {
         unsafe { &mut *self.inner.cast_mut() }.header.find_or_insert_group(key)
     }
+
+    /// Advance the collaborative-group cursor for `label` to the producer's
+    /// current write count, allocating the group slot if it does not exist.
+    /// Never rewinds (uses `fetch_max`).
+    ///
+    /// Use when joining a group whose backlog is meaningless — for instance,
+    /// when queued payloads index into an external arena that has since been
+    /// recycled, so replaying the ring would only surface stale references.
+    ///
+    /// MUST be called before any consumer in the group has begun consuming.
+    /// Calling it after some members have already claimed slots can race
+    /// with concurrent `acquire_next_slot` calls and silently lose any
+    /// messages the producer wrote in the fast-forward window. Typical
+    /// pattern: invoke once from the spine owner, immediately before
+    /// attaching the group's tiles.
+    pub fn fast_forward_collaborative_group(&self, label: &str) {
+        let key = format!("{}[{}].{}.collab", binary_name(), current_pid(), label);
+        let cursor = self.group_cursor(&key);
+        let head = self.count();
+        unsafe { (*cursor).fetch_max(head, Ordering::Relaxed) };
+    }
 }
 
 unsafe impl<T> Send for Queue<T> {}
diff --git a/crates/flux-communication/src/queue/tests_collaborative.rs b/crates/flux-communication/src/queue/tests_collaborative.rs
index 4b98d29..f969137 100644
--- a/crates/flux-communication/src/queue/tests_collaborative.rs
+++ b/crates/flux-communication/src/queue/tests_collaborative.rs
@@ -273,3 +273,71 @@ fn perf_test_collaborative_consumers() {
     assert_eq!(total_count, N, "received all {N} messages");
     println!("latency ({total_count} msgs) — mean: {}ns", total_sum / total_count as u64);
 }
+
+#[test]
+fn fast_forward_skips_backlog() {
+    const LABEL: &str = "ff_single";
+    let q: Queue<usize> = Queue::new(64, QueueType::MPMC);
+    let mut p = Producer::from(q);
+
+    for i in 0..30 {
+        p.produce(&i);
+    }
+
+    q.fast_forward_collaborative_group(LABEL);
+
+    let mut c = Consumer::new_collaborative_test(q, LABEL).without_log();
+
+    let mut received = Vec::new();
+    assert!(
+        !c.consume_collaborative(|x| received.push(*x)),
+        "post-fast-forward consumer must not see the backlog",
+    );
+    assert!(received.is_empty());
+
+    for i in 100..105 {
+        p.produce(&i);
+    }
+    while c.consume_collaborative(|x| received.push(*x)) {}
+    assert_eq!(received, vec![100, 101, 102, 103, 104]);
+}
+
+#[test]
+fn fast_forward_no_gaps_with_multiple_consumers() {
+    const LABEL: &str = "ff_multi";
+    const CONSUMERS: usize = 4;
+    const POST: usize = 16;
+
+    let q: Queue<usize> = Queue::new(64, QueueType::MPMC);
+    let mut p = Producer::from(q);
+
+    for i in 0..30 {
+        p.produce(&i);
+    }
+
+    q.fast_forward_collaborative_group(LABEL);
+
+    let mut consumers: Vec<_> =
+        (0..CONSUMERS).map(|_| Consumer::new_collaborative_test(q, LABEL).without_log()).collect();
+
+    for i in 100..(100 + POST) {
+        p.produce(&i);
+    }
+
+    let mut received = Vec::new();
+    for _ in 0..(POST * 4) {
+        for c in &mut consumers {
+            c.consume_collaborative(|x| received.push(*x));
+        }
+        if received.len() >= POST {
+            break;
+        }
+    }
+
+    received.sort_unstable();
+    assert_eq!(
+        received,
+        (100..(100 + POST)).collect::<Vec<_>>(),
+        "every post-fast-forward message must be delivered exactly once across the group",
+    );
+}
diff --git a/crates/flux/src/spine/mod.rs b/crates/flux/src/spine/mod.rs
index 9d3620a..a47b1fe 100644
--- a/crates/flux/src/spine/mod.rs
+++ b/crates/flux/src/spine/mod.rs
@@ -191,4 +191,19 @@ pub trait FluxSpine: Sized + Send {
         let _ = std::fs::remove_dir_all(shmem_dir(Self::app_name()))
             .inspect_err(|e| tracing::error!("couldn't remove spine queues {e}"));
     }
+
+    /// Fast-forward the collaborative group `label` for queue `T` to the
+    /// producer's current write head, dropping any backlog already in the
+    /// ring.
+    ///
+    /// Call once before attaching any tile in the group — see the
+    /// synchronisation note on
+    /// [`crate::communication::queue::Queue::fast_forward_collaborative_group`].
+    fn fast_forward_collaborative_group<T: 'static + Copy>(&self, label: &str)
+    where
+        Self: AsRef<SpineQueue<T>>,
+    {
+        let q: &SpineQueue<T> = self.as_ref();
+        q.fast_forward_collaborative_group(label);
+    }
 }