managedcode
diff --git a/‎docs/Features/AdminDashboard.md‎
Lines changed: 2 additions & 0 deletions b/‎docs/Features/AdminDashboard.md‎
Lines changed: 2 additions & 0 deletions
diff --git a/‎tests/KeyLoad.IntegrationTests/Features/AdminDashboard/AdminDashboardRf3Tests.cs‎
Lines changed: 4 additions & 1 deletion b/‎tests/KeyLoad.IntegrationTests/Features/AdminDashboard/AdminDashboardRf3Tests.cs‎
Lines changed: 4 additions & 1 deletion
@@ -114,3 +114,5 @@ The RF3 cases prove actual .NET/official MCP agreement, catalog/document/queue b
 The complete workflow concludes **failure**. Windows has four independent comparison-host startup timeouts: `AcHost003MissingSettingsReportFirstRequiredEndpoint`, `AcHost003InvalidDimensionsFailBeforeRequiredSettingsOrClients`, `AcHost003CommandLineOverridesInvalidEnvironmentDimension` and `AcHost003FirstEndpointAdvancesToSecondRequiredEndpoint`. Recovery is 119/121 on Linux/macOS and 118/121 on Windows; comparison-smoke also fails. These independently owned gates remain open. Passing RF3 does not establish endurance, power-loss durability or production readiness. Root numeric coverage thresholds remain mandatory, but no MTP/TUnit collector or baseline is configured, so no numeric coverage pass is claimed. No local tests were executed. ADR-051 stays Accepted until all required qualification is complete.
 
 Repair traceability: [run 37032546228](https://github.com/managedcode/KeyLoad/actions/runs/37032546228) exposed slash-equivalent routes and fixture epoch/time defects; [run 37036628601](https://github.com/managedcode/KeyLoad/actions/runs/37036628601) exposed empty query projection; [run 37038422388](https://github.com/managedcode/KeyLoad/actions/runs/37038422388) passed the browser but did not retain its images. The final source qualifies those repairs, native logging lifecycle, the independent 50-tool discovery oracle and canonical EventStreams expectations. Earlier compiler failures and cancelled runs are not passing evidence.
+
+Test correction (2026-10-02, exact run 37063262333): `AcAd007` compared the read-cut positions of an SDK read on node1 and an official MCP read on node3. Those are node-local applied positions of two independent gated reads, and concurrent RF3 writes separated them (92 vs 91). The test now requires both cuts to be real (positive). It keeps exact equality of the queue items, counters and catalog items, which is the caller-visible contract.
@@ -30,7 +30,10 @@ public async Task AcAd007RealSdkAndOfficialMcpAgreeOnCatalogAndNonconsumingQueue
             AdminDashboardProtocol.QueueTool, scenario.Messages, deadline.Token));
         await Assert.That(queue.Counters).IsEqualTo(queueMcp.Value.Counters);
         await Assert.That(queue.Items.SequenceEqual(queueMcp.Value.Items)).IsTrue();
-        await Assert.That(queue.CutPosition).IsEqualTo(queueMcp.Value.CutPosition);
+        // Cut positions are node-local applied positions of two independent gated reads on different physical nodes;
+        // concurrent RF3 writes legitimately separate them. The contract is identical persisted data at a real cut.
+        await Assert.That(queue.CutPosition).IsGreaterThan(0);
+        await Assert.That(queueMcp.Value.CutPosition).IsGreaterThan(0);
         await Assert.That(queue.Items[0].State).IsEqualTo(MessageState.Ready);
         await Assert.That(queue.Counters.StoredMessages).IsEqualTo(1);
         await Assert.That(catalogMcp.RequestId).IsNotEqualTo(queueMcp.RequestId);