diff --git a/crates/hypercolor-windows-capture/Cargo.toml b/crates/hypercolor-windows-capture/Cargo.toml
index 21fa19562..08ec19011 100644
--- a/crates/hypercolor-windows-capture/Cargo.toml
+++ b/crates/hypercolor-windows-capture/Cargo.toml
@@ -40,6 +40,7 @@ windows = { version = "0.62", default-features = false, features = [
   "Win32_Graphics_Gdi",
   "Win32_Security",
   "Win32_System_LibraryLoader",
+  "Win32_System_Threading",
 ] }
 
 [[bench]]
diff --git a/crates/hypercolor-windows-capture/src/duplication/gpu_surface.rs b/crates/hypercolor-windows-capture/src/duplication/gpu_surface.rs
index 0bbf6660e..f14ab3248 100644
--- a/crates/hypercolor-windows-capture/src/duplication/gpu_surface.rs
+++ b/crates/hypercolor-windows-capture/src/duplication/gpu_surface.rs
@@ -2609,6 +2609,45 @@ pub(super) mod fixture {
         publication.shared.texture_handle.borrowed().as_raw()
     }
 
+    pub(crate) fn wait_for_surface_progress(plan: &PreparedGpuSurfacePlan) -> CaptureResult<()> {
+        use windows::Win32::Foundation::WAIT_OBJECT_0;
+        use windows::Win32::System::Threading::{CreateEventW, WaitForSingleObject};
+
+        for slot in plan.routes.iter().flat_map(|route| &route.slots) {
+            let value = match slot.publication.state.load(Ordering::Acquire) {
+                USE_UNCLAIMED => slot.publication.synchronization.producer_ready_value,
+                USE_RELEASE_QUEUED => slot.publication.synchronization.consumer_release_value,
+                _ => continue,
+            };
+            // SAFETY: default security and no name create a private auto-reset event.
+            let event = OwnedHandle::new(
+                unsafe { CreateEventW(None, false, false, None) }
+                    .map_err(|error| CaptureError::windows("create fixture fence event", error))?,
+            )?;
+            // SAFETY: the plan owns the fence and event remains live through the wait.
+            unsafe {
+                slot.publication
+                    .shared
+                    .fence
+                    .SetEventOnCompletion(value, event.0)
+            }
+            .map_err(|error| CaptureError::windows("arm fixture fence event", error))?;
+            // SAFETY: event is a valid owned handle. The bound diagnoses a stalled
+            // native queue; it is not a publication cadence or polling budget.
+            let result = unsafe { WaitForSingleObject(event.0, 10_000) };
+            if result != WAIT_OBJECT_0 {
+                return Err(CaptureError::windows(
+                    "wait for fixture GPU progress",
+                    format!(
+                        "fence {value}, wait {result:?}, slots {:?}",
+                        slot_diagnostics(plan)
+                    ),
+                ));
+            }
+        }
+        Ok(())
+    }
+
     pub(crate) fn slot_diagnostics(
         plan: &PreparedGpuSurfacePlan,
     ) -> Vec<(usize, u8, u64, Option<u64>)> {
diff --git a/crates/hypercolor-windows-capture/src/duplication/tests.rs b/crates/hypercolor-windows-capture/src/duplication/tests.rs
index 47091569c..e8afe3977 100644
--- a/crates/hypercolor-windows-capture/src/duplication/tests.rs
+++ b/crates/hypercolor-windows-capture/src/duplication/tests.rs
@@ -894,7 +894,10 @@ fn abandoned_exact_gpu_surfaces_reclaim_under_sustained_pressure() {
     fixture.outcomes.clear();
 
     for sequence in 42..74 {
-        let publication = (0..64)
+        // A busy slot crosses at most two GPU boundaries: producer readiness,
+        // then the abandoned release. Observe those fences before retrying so
+        // concurrent WARP scheduling cannot consume an arbitrary polling budget.
+        let publication = (0..3)
             .find_map(|_| {
                 let outcomes = super::gpu_surface::fixture::republish(&mut fixture, sequence)
                     .expect("abandoned slots remain safely reclaimable");
@@ -903,7 +906,8 @@ fn abandoned_exact_gpu_surfaces_reclaim_under_sustained_pressure() {
                     crate::GpuSurfacePublishOutcome::Busy(_) => None,
                 });
                 if publication.is_none() {
-                    std::thread::sleep(std::time::Duration::from_millis(1));
+                    super::gpu_surface::fixture::wait_for_surface_progress(&fixture.plan)
+                        .expect("queued native work completes before retrying publication");
                 }
                 publication
             })