diff --git a/src/main.rs b/src/main.rs
index 0c6478c..1f24c22 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -2387,6 +2387,7 @@ async fn rfm_async_loop(
     let mut phase_st = OnlineStats::new("fresh_phase_us", 200, -1_500_000, 1_500_000);
     let mut phase_est: i64 = 0; // estimated TwinCAT frame boundary phase (ns within 1ms)
     let mut wphase_st = OnlineStats::new("write_phase_us", 200, -1_500_000, 1_500_000);
+    let mut j2cx_pending: Option<std::time::Instant> = None;
 
     while !STOP.load(Ordering::Relaxed) {
         if duration_sec > 0 && start.elapsed().as_secs() >= duration_sec {
@@ -2549,16 +2550,26 @@ async fn rfm_async_loop(
                     while (rt::now_ns() as i64) < target { core::hint::spin_loop(); }
                     wphase_st.push(((rt::now_ns() % 1_000_000) / 1000) as i64);
                 }
-                let tx_ok = ethercrab::Command::fpwr(station, SM_TX_ADDR)
-                    .send(maindevice, &slot[..PAYLOAD_LEN])
-                    .await
-                    .is_ok();
-                n_fpwr += 1;
-                if tx_ok {
-                    j2cx_st.push(t_detect.elapsed().as_nanos() as i64);
-                } else {
-                    n_fpwr_err += 1;
-                }
+                // PIPELINE v2: mirror the fprd's working pattern - a short
+                // yield-kick lets the executor run the TX task (frame goes on
+                // the wire), then store for poll-based completion at reap.
+                let mut f: FpwrFut<'_> = Box::pin(
+                    ethercrab::Command::fpwr(station, SM_TX_ADDR).send(maindevice, &slot[..PAYLOAD_LEN]),
+                );
+                let t_kick = Instant::now();
+                let _ = futures_lite::future::or(
+                    async { f.as_mut().await.is_ok() },
+                    async {
+                        while t_kick.elapsed().as_micros() < 20 {
+                            futures_lite::future::yield_now().await;
+                        }
+                        false
+                    },
+                )
+                .await;
+                pending_fpwr = Some(f);
+                j2cx_pending = Some(t_detect);
+                n_fpwr += 1; // issued; result counted at reap
             } else {
                 n_ovf += 1; // wire busy; newer tx_latest will go out after reap
             }
@@ -2635,6 +2646,9 @@ async fn rfm_async_loop(
                 if res.is_err() {
                     n_fpwr_err += 1;
                 }
+                if let Some(td) = j2cx_pending.take() {
+                    j2cx_st.push(td.elapsed().as_nanos() as i64);
+                }
             }
         }
         { let n = Instant::now(); seg_fpwr_us = n.duration_since(cursor).as_micros() as u64; cursor = n; }