fix: stabilize pull consumer expires timeout fetch

src/NATS.Server/JetStream/Consumers/PullConsumerEngine.cs

@@ -95,9 +95,6 @@ public sealed class CompiledFilter
 
 public sealed class PullConsumerEngine
 {
-    // Reusable fetch buffer to avoid per-fetch List allocation.
-    // Go reference: consumer.go — reuses slice backing array across fetch calls.
-    [ThreadStatic] private static List<StoredMessage>? t_fetchBuf;
     // Go: consumer.go — cluster-wide pending pull request tracking keyed by reply subject.
     // Reference: golang/nats-server/server/consumer.go waitingRequestsPending / proposeWaitingRequest
     private readonly ConcurrentDictionary<string, PullWaitingRequest> _clusterPending =
@@ -158,11 +155,7 @@ public sealed class PullConsumerEngine
     public async ValueTask<PullFetchBatch> FetchAsync(StreamHandle stream, ConsumerHandle consumer, PullFetchRequest request, CancellationToken ct)
     {
         var batch = Math.Max(request.Batch, 1);
-        // Use thread-static buffer to avoid per-fetch List allocation.
-        // Results are snapshot'd into PullFetchBatch before the buffer is reused.
-        var messages = t_fetchBuf ??= new List<StoredMessage>();
-        messages.Clear();
-        if (messages.Capacity < batch) messages.Capacity = batch;
+        var messages = new List<StoredMessage>(batch);
 
         // Go: consumer.go — enforce ExpiresMs timeout on pull fetch requests.
         // When ExpiresMs > 0, create a linked CancellationTokenSource that fires
@@ -313,6 +306,7 @@ public sealed class PullConsumerEngine
             await Task.Delay(5, ct).ConfigureAwait(false);
         }
 
+        ct.ThrowIfCancellationRequested();
         return null;
     }