diff --git a/src/DiffEngineViewer.Tests/ModuleInitializer.cs b/src/DiffEngineViewer.Tests/ModuleInitializer.cs
index 1dd21197..e9e334d1 100644
--- a/src/DiffEngineViewer.Tests/ModuleInitializer.cs
+++ b/src/DiffEngineViewer.Tests/ModuleInitializer.cs
@@ -11,5 +11,29 @@ public static void Initialize()
         // Before anything can touch DiffTools, which resolves and caches on first use. Only points
         // the resolver at the local build; nothing here launches anything.
         ManualViewer.Register();
+        RaiseThreadPoolFloor();
+    }
+
+    /// <summary>
+    /// The socket tests need two pool threads per exchange, and a CI runner starts with about four
+    /// in total.
+    /// <para>
+    /// <see cref="ServerFixture"/> drives the real <c>ViewerClient</c>, whose send is synchronous,
+    /// so a test blocks a pool thread for the whole exchange while the server needs threads of its
+    /// own to accept the connection and answer it. Run enough of those at once on a four core
+    /// runner and the answer waits on the pool's hill climb, which adds roughly one thread a
+    /// second, past the client's three second timeout. That surfaced as an intermittent "No
+    /// response for Inline", on a different platform and a different test each run.
+    /// </para>
+    /// <para>
+    /// Raising the floor removes the scarcity rather than the blocking. The blocking is real, but
+    /// it belongs to the client the tray and an attached viewer actually use, and no process runs
+    /// dozens of those at once — a test host is the only thing that does.
+    /// </para>
+    /// </summary>
+    static void RaiseThreadPoolFloor()
+    {
+        ThreadPool.GetMinThreads(out var workers, out var completionPorts);
+        ThreadPool.SetMinThreads(Math.Max(workers, 32), completionPorts);
     }
 }