test(ffi): de-flake pump_yields_when_queue_full via threshold recalibration

sdks/c/src/event_queue.rs

@@ -752,14 +752,17 @@ mod phase2_regression_tests {
         let canary_after = canary.load(Ordering::Relaxed);
         canary_stop.store(1, Ordering::Relaxed);
 
-        // The drain takes >=NUM_EVENTS*10ms = 1000ms. The canary should
-        // have made many ticks during that time IF the producer yielded.
-        // If push_event_with_capacity busy-spinned, the single worker
-        // would have been monopolized and the canary would barely advance.
+        // `boxlite_runtime_drain(rt, 0, …)` empties the queue in bulk, so the
+        // drainer finishes in ~150-250ms — not the ~1000ms a naive
+        // "100 events × 10ms" estimate suggests. In that window a cooperative
+        // producer lets the canary tick ~8-10 times; a producer that busy-spins
+        // or blocks the single worker instead collapses it to ~2 (measured).
+        // `>= 5` separates the two with margin on both sides, without sitting on
+        // the flaky 8-10 boundary the old `>= 10` threshold did.
         let progress = canary_after.saturating_sub(canary_before);
         assert!(
-            progress >= 10,
-            "canary advanced only {progress} ticks; producer likely busy-spinning instead of yielding"
+            progress >= 5,
+            "canary advanced only {progress} ticks; producer likely busy-spinning or blocking the worker instead of yielding"
         );
 
         unsafe { boxlite_runtime_free(rt_ptr) };