Fix - run CI tests on windows-latest as well as ubuntu

.github/workflows/ci.yml

@@ -10,7 +10,12 @@ permissions:
 
 jobs:
   build-test:
-    runs-on: ubuntu-latest
+    strategy:
+      # Let both legs finish so a Windows-only failure doesn't hide the Linux result (or the reverse).
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+    runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:

tests/EDNexus.App.Tests/PluginServiceExitTests.cs

@@ -148,7 +148,7 @@ public async Task ShutdownForExit_WhenAHandlerNeverReturns_GivesUpAtTheBound_And
 
         Assert.False(finishedInTime);
         Assert.False(latches.HandlerReturned.IsSet);
-        Assert.Null(service.Runtime);
+        Assert.True(SpinWait.SpinUntil(() => service.Runtime is null, Ceiling));
         Assert.Empty(_errors);
 
         latches.HandlerGate.Set();
@@ -166,7 +166,7 @@ public async Task Detach_WithABound_WhenAHandlerNeverReturns_ReturnsFalse_AndThe
 
         Assert.False(Bounded(() => service.Detach(ShortBound), "Detach(timeout)"));
 
-        Assert.Null(service.Runtime);
+        Assert.True(SpinWait.SpinUntil(() => service.Runtime is null, Ceiling));
         latches.HandlerGate.Set();
         Assert.True(latches.HandlerReturned.Wait(Ceiling));
     }

tests/EDNexus.App.Tests/PluginServiceLifecycleTests.cs

@@ -159,11 +159,14 @@ public async Task ASecondDetach_Joins_AndReturnsOnceTheFirstHasFinished()
         service.Attach(bus, state);
         await service.StartAsync();
 
-        var first = Task.Run(service.Detach);
+        // Dedicated threads and a blocking probe: both detaches block, and with the pool starved (seen on
+        // the Windows runner) a WaitAsync timeout can't fire until the host's 5 s shutdown bound has let
+        // the first one finish, so the "not done yet" check would pass or fail on scheduling, not on the join.
+        var first = Task.Factory.StartNew(() => service.Detach(), TaskCreationOptions.LongRunning);
         Assert.True(latches.Entered.Wait(Ceiling));
-        var second = Task.Run(() => service.Detach(Ceiling));
+        var second = Task.Factory.StartNew(() => service.Detach(Ceiling), TaskCreationOptions.LongRunning);
         // Joined to the first, which is parked: it cannot be done yet.
-        await Assert.ThrowsAsync<TimeoutException>(() => second.WaitAsync(TimeSpan.FromMilliseconds(200)));
+        Assert.False(SpinWait.SpinUntil(() => second.IsCompleted, TimeSpan.FromMilliseconds(200)));
 
         latches.Gate.Set();
 

tests/EDNexus.App.Tests/PluginServiceReattachTests.cs

@@ -125,7 +125,7 @@ public void Shutdown() { {{latches.Set("returned")}} }
 
         Assert.False(Bounded(() => service.Detach(ShortBound), "Detach(timeout)"));
 
-        Assert.Null(service.Runtime);               // detached, but its teardown is still waiting for the load pass
+        Assert.True(SpinWait.SpinUntil(() => service.Runtime is null, Ceiling));   // detached, but its teardown is still waiting for the load pass
         Assert.False(stuck.Gate.IsSet);
         service.Enablement.SetDisabled(stuckId, true);
         return (service, stuck);
@@ -154,7 +154,9 @@ public async Task Attach_WhileAnOlderTeardownIsStillRunning_DoesNotWaitForIt_And
         var finishedInTime = Bounded(() => service.ShutdownForExit(ShortBound), "ShutdownForExit");
 
         Assert.False(finishedInTime);              // the older teardown is still running
-        Assert.True(fresh.Returned.IsSet);         // but the new plugin's Shutdown ran: it is not left running at exit
+        // But the new plugin's Shutdown runs: it is not left running at exit. ShutdownForExit bounds the
+        // whole teardown, so on a slow machine it can return first; wait for it rather than assume it.
+        Assert.True(fresh.Returned.Wait(Ceiling));
         Assert.Null(service.Runtime);
         var (bus3, state3) = Engine();
         Assert.False(service.Attach(bus3, state3));   // closed for good

tests/EDNexus.Plugins.Hosting.Tests/PluginContentBindingTests.cs

@@ -318,8 +318,10 @@ private static string Hash(string dir, string? entry = null)
     [Fact]
     public void FolderHash_IsDeterministic_AndIndependentOfCreationOrder()
     {
-        var a = Tree("order-a", ("b.txt", "2"), ("a.txt", "1"), ("sub/z.txt", "3"), ("sub/y.txt", "4"), ("A.txt", "5"));
-        var b = Tree("order-b", ("A.txt", "5"), ("sub/y.txt", "4"), ("sub/z.txt", "3"), ("a.txt", "1"), ("b.txt", "2"));
+        // "C.txt" (not "A.txt"): an upper-case name that ordinal order puts first, without colliding
+        // with "a.txt" on a case-insensitive file system (Windows, macOS), where the two are one file.
+        var a = Tree("order-a", ("b.txt", "2"), ("a.txt", "1"), ("sub/z.txt", "3"), ("sub/y.txt", "4"), ("C.txt", "5"));
+        var b = Tree("order-b", ("C.txt", "5"), ("sub/y.txt", "4"), ("sub/z.txt", "3"), ("a.txt", "1"), ("b.txt", "2"));
 
         Assert.Equal(Hash(a), Hash(a));
         Assert.Equal(Hash(a), Hash(b));

tests/EDNexus.Tests/Twitch/TwitchStreamCardServiceTests.cs

@@ -772,6 +772,10 @@ public async Task A_clear_for_a_closed_game_that_fails_is_queued_and_dropped_onc
         Assert.Equal(EbsCleanupKind.ClearCard, pending.Kind);
         Assert.Equal("ebs-token", pending.Token);
 
+        // Stop the service first: its pump keeps retrying the failed clear (up to MaxPublishRetries), and
+        // each retry re-enqueues the entry while its own clear is in flight, so a count taken during one
+        // would see it. What is under test here is the queue's retry alone, as after an app restart.
+        service.Dispose();
         client.RespondToClear = () => StreamStatePublishResult.ClearedOk;
         Assert.Equal(0, await env.Queue.RetryPendingAsync());
         Assert.Empty(env.Store.Load().Twitch.PendingCleanups);