fix(worker): StaWaitHelper regression tests + full-drain contract notes; consolidate wait P/Invoke

src/ZB.MOM.WW.MxGateway.Worker/Sta/StaWaitHelper.cs

@@ -1,4 +1,5 @@
 using System;
+using System.Diagnostics;
 using System.Runtime.InteropServices;
 using System.Threading;
 using Microsoft.Win32.SafeHandles;
@@ -17,23 +18,22 @@ namespace ZB.MOM.WW.MxGateway.Worker.Sta;
 ///     Deliberately separate from <see cref="StaMessagePump"/> rather than a new
 ///     method on it: the two caches that use this are plain per-session objects
 ///     with no pump dependency (they take the pump step as an injected delegate
-///     so tests can drive them off the STA). Keeping the helper in this folder
-///     keeps every <c>user32</c> wait signature in the worker reviewable in one
-///     place; the declaration below mirrors the pump's exactly.
+///     so tests can drive them off the STA). The wait itself is not duplicated —
+///     both this and the pump call the single declaration in
+///     <see cref="StaNativeMethods"/>.
 /// </remarks>
 internal static class StaWaitHelper
 {
     /// <summary>
-    ///     Ceiling on the blind fallback tick. A caller's poll interval is
-    ///     clamped to this so no wait can go longer than 50 ms without calling
-    ///     its pump step, even when the process has no message queue to wake it
-    ///     (unit-test threads) and the wait handle is never signaled.
+    ///     Ceiling on the fallback tick — the longest a wait may block when
+    ///     nothing wakes it. A caller's poll interval is clamped to this, so a
+    ///     caller's pump step keeps running at least this often no matter what
+    ///     interval it asked for, even on a thread that never receives a message
+    ///     and whose signal is never set.
     /// </summary>
     internal const int MaxFallbackTickMilliseconds = 50;
 
-    private const uint MsgWaitFailed = 0xFFFFFFFF;
-    private const uint MwmoInputAvailable = 0x0004;
-    private const uint QsAllInput = 0x04FF;
+    private static int waitFailureReported;
 
     /// <summary>
     ///     Blocks the calling thread until <paramref name="signal"/> is set, a
@@ -42,6 +42,29 @@ internal static class StaWaitHelper
     ///     never blinds the thread to inbound COM event messages: the wait
     ///     returns the instant one arrives, so the caller can dispatch it.
     /// </summary>
+    /// <remarks>
+    ///     <para>
+    ///         <em>Caller contract: the pump step must drain the message queue
+    ///         to empty on every call.</em> The wait uses
+    ///         <see cref="StaNativeMethods.MwmoInputAvailable"/>, which reports
+    ///         input that is merely <em>present</em> — including input an earlier
+    ///         <c>PeekMessage</c> has already seen but not removed. A caller
+    ///         whose pump step drains only part of the queue, or does nothing at
+    ///         all, therefore leaves a message that satisfies the wake condition
+    ///         forever: every wait returns immediately, and the caller's poll
+    ///         loop degenerates into a busy loop that spins the thread at 100%
+    ///         until its deadline. That is a CPU fault, not a correctness one —
+    ///         the loop still honors its deadline and return shape — but it must
+    ///         not ship. Drain with <c>PeekMessage(..., PM_REMOVE)</c> in a loop
+    ///         until it returns false, exactly as
+    ///         <see cref="StaMessagePump.PumpPendingMessages"/> does.
+    ///     </para>
+    ///     <para>
+    ///         The wait does not <c>DangerousAddRef</c> the handle. That is
+    ///         sound only because the callers never dispose their wait handles —
+    ///         see the coupled invariant documented on the caches' signal fields.
+    ///     </para>
+    /// </remarks>
     /// <param name="signal">Wait handle set by whoever produces the awaited result.</param>
     /// <param name="timeoutMilliseconds">Maximum wait; values at or below zero return immediately.</param>
     internal static void WaitForSignalOrMessages(
@@ -60,20 +83,28 @@ internal static class StaWaitHelper
 
         SafeWaitHandle safeHandle = signal.SafeWaitHandle;
         IntPtr[] handles = [safeHandle.DangerousGetHandle()];
-        uint result = MsgWaitForMultipleObjectsEx(
+        uint result = StaNativeMethods.MsgWaitForMultipleObjectsEx(
             (uint)handles.Length,
             handles,
             (uint)timeoutMilliseconds,
-            QsAllInput,
-            MwmoInputAvailable);
+            StaNativeMethods.QsAllInput,
+            StaNativeMethods.MwmoInputAvailable);
+
+        // Read before anything else can run another P/Invoke and overwrite the
+        // thread's last-error slot.
+        int lastError = result == StaNativeMethods.MsgWaitFailed
+            ? Marshal.GetLastWin32Error()
+            : 0;
 
         // The handle is a field of a live caller object, so it cannot be
         // collected mid-call; KeepAlive states that dependency explicitly
         // rather than relying on it.
         GC.KeepAlive(signal);
 
-        if (result == MsgWaitFailed)
+        if (result == StaNativeMethods.MsgWaitFailed)
         {
+            ReportWaitFailureOnce(lastError);
+
             // Degrade to the pre-existing blind sleep instead of throwing or
             // hot-spinning: the caller is a bounded, deadline-driven poll loop,
             // and a failed wait must not turn a timed-out write into a fault.
@@ -108,11 +139,30 @@ internal static class StaWaitHelper
             : tick;
     }
 
-    [DllImport("user32.dll", SetLastError = true)]
-    private static extern uint MsgWaitForMultipleObjectsEx(
-        uint count,
-        IntPtr[] handles,
-        uint milliseconds,
-        uint wakeMask,
-        uint flags);
+    /// <summary>
+    ///     Traces the first wait failure of the process and stays silent after
+    ///     that.
+    /// </summary>
+    /// <remarks>
+    ///     The degrade path is silent by design — a failed wait still honors the
+    ///     caller's deadline — but a <em>permanently</em> failing wait means every
+    ///     completion wait has quietly reverted to the sleep-polled behavior this
+    ///     class exists to replace, and nothing else would ever say so. One trace
+    ///     makes that discoverable; latching it keeps a repeating failure from
+    ///     flooding the trace listener from inside a poll loop.
+    /// </remarks>
+    /// <param name="lastError">The Win32 error captured immediately after the failed wait.</param>
+    private static void ReportWaitFailureOnce(int lastError)
+    {
+        if (Interlocked.Exchange(ref waitFailureReported, 1) != 0)
+        {
+            return;
+        }
+
+        Trace.TraceWarning(
+            "MsgWaitForMultipleObjectsEx failed in the worker STA completion wait (Win32 error {0}). "
+            + "The wait has degraded to blind sleeping: deadlines are still honored, but inbound "
+            + "MXAccess COM events will only dispatch at poll-tick granularity. Reported once per process.",
+            lastError);
+    }
 }