perf(ui): shared KPI cache, live-cache-backed alarm summary, coalesced debug renders

src/ZB.MOM.WW.ScadaBridge.CentralUI/Components/Pages/Deployment/DebugView.razor

@@ -291,13 +291,56 @@
                 .OrderBy(a => a.AttributeName)
                 .ToList();
 
-    /// <summary>Attribute composition forest, rebuilt from the live latest-per-name dictionary.</summary>
+    // ── Tree memoization (arch-review WP2.4) ──────────────────────────────────
-    private IReadOnlyList<DebugTreeNode> AttributeForest =>
+    // Both forests used to be rebuilt from scratch on EVERY render — and a render
-        DebugTreeBuilder.BuildAttributeTree(_attributeValues.Values, _attrFilter);
+    // fired per streamed event. A busy instance therefore rebuilt two whole
+    // composition trees hundreds of times a second, on the circuit's single
+    // dispatcher thread. The rebuild now sits behind a version stamp bumped
+    // whenever the underlying dictionaries change (a coalesced stream flush,
+    // Connect, Disconnect), so an unrelated re-render (tab switch, toast, status
+    // strip tick) reuses the forest instead of rebuilding it. Reusing the same
+    // instance also lets Blazor skip re-rendering the TreeView subtree entirely.
+    //
+    // The cached key includes the dictionary Count as well as the version: the
+    // trees must never go stale if the dictionaries are populated by some path
+    // that forgets to bump the stamp.
+    private int _dataVersion;
 
-    /// <summary>Alarm composition forest, rebuilt from the live latest-per-name dictionary.</summary>
+    private IReadOnlyList<DebugTreeNode>? _attrForestCache;
-    private IReadOnlyList<DebugTreeNode> AlarmForest =>
+    private (int Version, int Count, string Filter) _attrForestKey = (-1, -1, string.Empty);
-        DebugTreeBuilder.BuildAlarmTree(_alarmStates.Values, _alarmFilter);
+
+    private IReadOnlyList<DebugTreeNode>? _alarmForestCache;
+    private (int Version, int Count, string Filter) _alarmForestKey = (-1, -1, string.Empty);
+
+    /// <summary>Attribute composition forest, rebuilt only when the underlying data or filter changed.</summary>
+    private IReadOnlyList<DebugTreeNode> AttributeForest
+    {
+        get
+        {
+            var key = (_dataVersion, _attributeValues.Count, _attrFilter);
+            if (_attrForestCache is null || _attrForestKey != key)
+            {
+                _attrForestCache = DebugTreeBuilder.BuildAttributeTree(_attributeValues.Values, _attrFilter);
+                _attrForestKey = key;
+            }
+            return _attrForestCache;
+        }
+    }
+
+    /// <summary>Alarm composition forest, rebuilt only when the underlying data or filter changed.</summary>
+    private IReadOnlyList<DebugTreeNode> AlarmForest
+    {
+        get
+        {
+            var key = (_dataVersion, _alarmStates.Count, _alarmFilter);
+            if (_alarmForestCache is null || _alarmForestKey != key)
+            {
+                _alarmForestCache = DebugTreeBuilder.BuildAlarmTree(_alarmStates.Values, _alarmFilter);
+                _alarmForestKey = key;
+            }
+            return _alarmForestCache;
+        }
+    }
 
     private DebugStreamSession? _session;
     private ToastNotification _toast = default!;
@@ -458,6 +501,9 @@
             foreach (var al in session.InitialSnapshot.AlarmStates)
                 _alarmStates[al.AlarmName] = al;
 
+            // Snapshot seeding is a data change like any other — bump the stamp so
+            // the memoized forests rebuild from it.
+            _dataVersion++;
             _snapshot = session.InitialSnapshot;
             _connected = true;
 
@@ -492,6 +538,11 @@
         _snapshot = null;
         _attributeValues.Clear();
         _alarmStates.Clear();
+        // Drop anything the stream buffered for the session we just tore down, so a
+        // late flush can't repopulate a disconnected view.
+        _pendingAttributes.Clear();
+        _pendingAlarms.Clear();
+        _dataVersion++;
     }
 
     /// <summary>
@@ -508,37 +559,108 @@
         _toast.ShowInfo("Cleared previous session — select a site and instance to begin.", autoDismissMs: 5000);
     }
 
+    // ── Render coalescing (arch-review WP2.4) ─────────────────────────────────
+    // Every streamed event used to marshal onto the dispatcher and call
+    // StateHasChanged individually, so a chatty instance drove one full render
+    // (and two full tree rebuilds) per value change. Events now land in a
+    // thread-safe pending map keyed by name — repeated updates to the same tag
+    // inside one window collapse to the latest — and exactly ONE dispatcher
+    // marshal per window drains them, bumps the version stamp and renders once.
+    private const int RenderCoalesceMs = 250;
+
+    private readonly System.Collections.Concurrent.ConcurrentDictionary<string, AttributeValueChanged> _pendingAttributes = new();
+    private readonly System.Collections.Concurrent.ConcurrentDictionary<string, AlarmStateChanged> _pendingAlarms = new();
+
+    /// <summary>0/1 flag: whether a coalesce window is already armed. Interlocked-guarded.</summary>
+    private int _flushArmed;
+
     /// <summary>
     /// Handles one debug-stream event. The callback is invoked on an Akka/gRPC
     /// thread, but <see cref="_attributeValues"/>/<see cref="_alarmStates"/> are
     /// <see cref="Dictionary{TKey,TValue}"/> instances also enumerated by the
     /// render thread (the tree forests + <see cref="FilteredAttributeValues"/> are
     /// built from them). <c>Dictionary</c> is not thread-safe (CentralUI-021): a
-    /// write racing an enumeration can throw or corrupt the buckets. The mutation
+    /// write racing an enumeration can throw or corrupt the buckets. The event is
-    /// is therefore marshalled onto the renderer's dispatcher via
+    /// therefore parked in a concurrent pending map here and applied on the
-    /// <see cref="SafeInvokeAsync"/> so every access to the dictionaries — read
+    /// renderer's dispatcher by <see cref="FlushPendingAsync"/>, so every access to
-    /// and write — happens on the render thread.
+    /// the render dictionaries — read and write — still happens on one thread.
     /// </summary>
     private void HandleStreamEvent(object evt)
     {
         // CentralUI-009: the component may have been disposed while this event
         // was in flight on the Akka/gRPC thread.
         if (_disposed) return;
-        _ = SafeInvokeAsync(() =>
+
-        {
-            if (_disposed) return;
         switch (evt)
         {
             case AttributeValueChanged av:
-                    _attributeValues[av.AttributeName] = av;
+                _pendingAttributes[av.AttributeName] = av;
                 break;
             case AlarmStateChanged al:
-                    _alarmStates[al.AlarmName] = al;
+                _pendingAlarms[al.AlarmName] = al;
                 break;
             default:
-                    // Unknown event type — no re-render needed.
+                // Unknown event type — nothing to apply, no render needed.
                 return;
         }
+
+        // Arm the window if it isn't already. Losers of the CAS have nothing to do:
+        // their event is in the pending map and the armed flush will pick it up.
+        if (Interlocked.CompareExchange(ref _flushArmed, 1, 0) == 0)
+        {
+            _ = FlushPendingAsync();
+        }
+    }
+
+    /// <summary>
+    /// Waits out the coalesce window, then applies everything buffered in one
+    /// dispatcher pass and renders once.
+    /// </summary>
+    private async Task FlushPendingAsync()
+    {
+        try
+        {
+            await Task.Delay(RenderCoalesceMs);
+        }
+        catch
+        {
+            // Delay never faults in practice; if it ever did, fall through and drain.
+        }
+
+        if (_disposed)
+        {
+            Volatile.Write(ref _flushArmed, 0);
+            return;
+        }
+
+        await SafeInvokeAsync(() =>
+        {
+            // Disarm BEFORE draining: an event arriving mid-drain then arms a fresh
+            // window rather than being stranded until the next unrelated event.
+            Volatile.Write(ref _flushArmed, 0);
+            if (_disposed) return;
+
+            var applied = false;
+            foreach (var key in _pendingAttributes.Keys)
+            {
+                if (_pendingAttributes.TryRemove(key, out var av))
+                {
+                    _attributeValues[av.AttributeName] = av;
+                    applied = true;
+                }
+            }
+            foreach (var key in _pendingAlarms.Keys)
+            {
+                if (_pendingAlarms.TryRemove(key, out var al))
+                {
+                    _alarmStates[al.AlarmName] = al;
+                    applied = true;
+                }
+            }
+
+            if (!applied) return;
+
+            _dataVersion++;
             StateHasChanged();
         });
     }

src/ZB.MOM.WW.ScadaBridge.CentralUI/Components/Pages/Deployment/Deployments.razor

@@ -236,13 +236,66 @@
         DeploymentStatusNotifier.StatusChanged += OnDeploymentStatusChanged;
     }
 
+    // ── Push coalescing (arch-review WP2.4) ───────────────────────────────────
+    // One reload = every deployment record + every instance, re-read and re-filtered.
+    // The notifier fires per status WRITE, so a site deploy of N instances produced
+    // 2N+ of those full reloads back to back, per circuit. The reload is now
+    // leading-edge debounced: the first push after an idle gap still reloads
+    // immediately (a single deployment stays as responsive as before), and every
+    // push inside the window is absorbed into ONE trailing reload.
+    private const int ReloadDebounceMs = 500;
+
+    private DateTimeOffset _lastReloadAt = DateTimeOffset.MinValue;
+
+    /// <summary>Coalescing timer for pushes arriving inside the debounce window. Disposed with the component.</summary>
+    private Timer? _coalesceTimer;
+
     private void OnDeploymentStatusChanged(ZB.MOM.WW.ScadaBridge.DeploymentManager.DeploymentStatusChange change)
     {
         // CentralUI-022: a callback racing disposal must not touch the component.
         if (_disposed || !_autoRefresh) return;
+
+        lock (_coalesceLock)
+        {
+            if (_disposed) return;
+
+            var sinceLast = DateTimeOffset.UtcNow - _lastReloadAt;
+            if (sinceLast >= TimeSpan.FromMilliseconds(ReloadDebounceMs))
+            {
+                // Idle — reload straight away (leading edge).
+                _lastReloadAt = DateTimeOffset.UtcNow;
+            }
+            else
+            {
+                // Inside the window: arm one trailing reload for the remainder, and
+                // let any further push in this window ride it.
+                if (_coalesceTimer is not null) return;
+
+                var delay = TimeSpan.FromMilliseconds(ReloadDebounceMs) - sinceLast;
+                _coalesceTimer = new Timer(_ => OnCoalesceElapsed(), null, delay, Timeout.InfiniteTimeSpan);
+                return;
+            }
+        }
+
         _ = DispatchReloadAsync();
     }
 
+    /// <summary>Trailing edge of the debounce: disarm, stamp, and run the one coalesced reload.</summary>
+    private void OnCoalesceElapsed()
+    {
+        lock (_coalesceLock)
+        {
+            _coalesceTimer?.Dispose();
+            _coalesceTimer = null;
+            if (_disposed) return;
+            _lastReloadAt = DateTimeOffset.UtcNow;
+        }
+
+        _ = DispatchReloadAsync();
+    }
+
+    private readonly object _coalesceLock = new();
+
     /// <summary>
     /// Reloads the deployment table on the renderer's dispatcher, guarded
     /// against the component being disposed mid-flight (CentralUI-022):
@@ -361,5 +414,10 @@
         // status change reaches this disposed component.
         _disposed = true;
         DeploymentStatusNotifier.StatusChanged -= OnDeploymentStatusChanged;
+        lock (_coalesceLock)
+        {
+            _coalesceTimer?.Dispose();
+            _coalesceTimer = null;
+        }
     }
 }

src/ZB.MOM.WW.ScadaBridge.CentralUI/Components/Pages/Monitoring/AlarmSummary.razor

@@ -1,5 +1,6 @@
 @page "/monitoring/alarms"
 @attribute [Authorize(Policy = ZB.MOM.WW.ScadaBridge.Security.AuthorizationPolicies.RequireDeployment)]
+@using Microsoft.AspNetCore.Components.Web.Virtualization
 @using ZB.MOM.WW.ScadaBridge.CentralUI.Components.Shared
 @using ZB.MOM.WW.ScadaBridge.CentralUI.Services
 @using ZB.MOM.WW.ScadaBridge.Commons.Entities.Sites
@@ -196,15 +197,25 @@
                                 @onkeydown='e => OnHeaderKeyDown(e, "severity")' @onkeydown:preventDefault="_preventHeaderDefault">Severity @SortGlyph("severity")</th>
                         </tr>
                     </thead>
+                    @* Row markup is identical on both paths — the same <tr>, the same cells,
+                       the same data-test hook. Only WHO enumerates it differs: a plain
+                       foreach for the ordinary case, and Virtualize once the flat table
+                       gets long enough that rendering every row per delta dominates the
+                       circuit (arch-review WP2.4). Virtualize needs <tr> spacers or the
+                       browser hoists its <div>s out of the table and the layout collapses. *@
                     <tbody>
+                        @if (_visibleRows.Count <= VirtualizeThreshold)
+                        {
                             @foreach (var row in _visibleRows)
                             {
-                            <tr data-test="alarm-summary-row">
+                                @AlarmRow(row)
-                                <td class="font-monospace small">@row.InstanceUniqueName</td>
+                            }
-                                <td>@row.Alarm.AlarmName</td>
+                        }
-                                <td><AlarmStateBadges Alarm="row.Alarm" /></td>
+                        else
-                                <td class="text-end font-monospace">@row.Alarm.Condition.Severity</td>
+                        {
-                            </tr>
+                            <Virtualize Items="_visibleRows" Context="row" ItemSize="37" SpacerElement="tr">
+                                @AlarmRow(row)
+                            </Virtualize>
                         }
                     </tbody>
                 </table>
@@ -222,7 +233,23 @@
     // P4 (arch-review): the filtered+sorted view is memoized rather than recomputed
     // on every render. RecomputeVisibleRows() refreshes it whenever the inputs change
     // (a fresh snapshot in RefreshAsync, a filter change via @bind:after, or a sort).
-    private IReadOnlyList<AlarmSummaryRow> _visibleRows = Array.Empty<AlarmSummaryRow>();
+    // Concrete List because Virtualize binds ICollection<T>, not IReadOnlyList<T>.
+    private List<AlarmSummaryRow> _visibleRows = new();
+
+    // Above this many visible rows the flat table switches to Virtualize. Below it the
+    // plain foreach is cheaper than the virtualization machinery (and needs no JS), and
+    // an operator filtered down to a handful of alarms should never pay for either.
+    private const int VirtualizeThreshold = 150;
+
+    // One row template, shared by the plain and virtualized paths so the two can never
+    // drift visually.
+    private RenderFragment<AlarmSummaryRow> AlarmRow => row =>
+        @<tr data-test="alarm-summary-row">
+            <td class="font-monospace small">@row.InstanceUniqueName</td>
+            <td>@row.Alarm.AlarmName</td>
+            <td><AlarmStateBadges Alarm="row.Alarm" /></td>
+            <td class="text-end font-monospace">@row.Alarm.Condition.Severity</td>
+        </tr>;
 
     private IReadOnlyList<string> _notReporting = Array.Empty<string>();
     private AlarmRollup _rollup = new(0, 0, 0, new Dictionary<AlarmKind, int>());

src/ZB.MOM.WW.ScadaBridge.CentralUI/Components/Pages/Monitoring/Health.razor

@@ -15,8 +15,7 @@
 @implements IDisposable
 @inject ICentralHealthAggregator HealthAggregator
 @inject ISiteRepository SiteRepository
-@inject CommunicationService CommunicationService
+@inject IKpiSnapshotCache KpiSnapshots
-@inject IAuditLogQueryService AuditLogQueryService
 @inject IKpiHistoryQueryService KpiHistory
 @inject Microsoft.Extensions.Options.IOptions<HealthMonitoringOptions> HealthOptions
 
@@ -640,7 +639,9 @@
             // Non-fatal — fall back to showing siteId only
         }
 
-        await RefreshNow();
+        // First load rides whatever the shared KPI cache already has fresh — a second
+        // operator opening the dashboard inside the window costs no extra query.
+        await RefreshAllAsync(forceRefresh: false);
 
         // Site Health Trends (M6) load on their own path — never from the
         // timer tick below — so a trend-query fault can't disturb the live tile
@@ -655,7 +656,7 @@
             {
                 try
                 {
-                    await RefreshNow();
+                    await RefreshAllAsync(forceRefresh: false);
                     StateHasChanged();
                 }
                 finally
@@ -769,20 +770,38 @@
             ? "Central Cluster"
             : $"{GetSiteName(siteKey)} ({siteKey})";
 
-    private async Task RefreshNow()
+    // Per-tile budget for one refresh tick. The underlying KPI Asks carry the
+    // 30s CommunicationOptions.QueryTimeout, so without this a single hung
+    // singleton stalled the WHOLE tick (and, with a 10s timer, stacked ticks)
+    // rather than degrading its own tile. Cancelling here abandons only this
+    // circuit's wait — the shared KPI flight continues and its result lands in
+    // the process-level cache for the next tick.
+    private static readonly TimeSpan TileTimeout = TimeSpan.FromSeconds(5);
+
+    // The operator's explicit "Refresh Now" bypasses the shared cache's freshness
+    // window; the timer tick and the first load ride whatever is already fresh
+    // (which is the whole point — N dashboards share one query per window).
+    private Task RefreshNow() => RefreshAllAsync(forceRefresh: true);
+
+    // Every tile loads concurrently, each with its own short budget, so one slow
+    // KPI degrades one tile instead of the tick. Every loader swallows its own
+    // faults, so Task.WhenAll here can never observe an exception.
+    private async Task RefreshAllAsync(bool forceRefresh)
     {
         _siteStates = HealthAggregator.GetAllSiteStates();
-        await LoadOutboxKpis();
+        await Task.WhenAll(
-        await Task.WhenAll(LoadSiteCallKpis(), LoadSiteCallNodeKpis());
+            LoadOutboxKpis(forceRefresh),
-        await LoadAuditKpis();
+            LoadSiteCallKpis(forceRefresh),
+            LoadSiteCallNodeKpis(forceRefresh),
+            LoadAuditKpis(forceRefresh));
     }
 
-    private async Task LoadOutboxKpis()
+    private async Task LoadOutboxKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetNotificationKpisAsync(
+            using var timeout = new CancellationTokenSource(TileTimeout);
-                new NotificationKpiRequest(Guid.NewGuid().ToString("N")));
+            var response = await KpiSnapshots.GetNotificationKpisAsync(forceRefresh, timeout.Token);
             if (response.Success)
             {
                 _outboxKpi = response;
@@ -795,6 +814,11 @@
                 _outboxKpiError = response.ErrorMessage ?? "KPI query failed.";
             }
         }
+        catch (OperationCanceledException)
+        {
+            _outboxKpiAvailable = false;
+            _outboxKpiError = "KPI query timed out.";
+        }
         catch (Exception ex)
         {
             _outboxKpiAvailable = false;
@@ -807,12 +831,12 @@
     // killing the dashboard. Mirrors LoadOutboxKpis's error handling shape — a
     // response with Success == false (repository fault) and an Ask that threw
     // (transport fault) both collapse to "unavailable".
-    private async Task LoadSiteCallKpis()
+    private async Task LoadSiteCallKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetSiteCallKpisAsync(
+            using var timeout = new CancellationTokenSource(TileTimeout);
-                new SiteCallKpiRequest(Guid.NewGuid().ToString("N")));
+            var response = await KpiSnapshots.GetSiteCallKpisAsync(forceRefresh, timeout.Token);
             if (response.Success)
             {
                 _siteCallKpi = response;
@@ -825,6 +849,11 @@
                 _siteCallKpiError = response.ErrorMessage ?? "KPI query failed.";
             }
         }
+        catch (OperationCanceledException)
+        {
+            _siteCallKpiAvailable = false;
+            _siteCallKpiError = "KPI query timed out.";
+        }
         catch (Exception ex)
         {
             _siteCallKpiAvailable = false;
@@ -834,12 +863,12 @@
 
     // Per-node site-call KPI loader (M5.2). Best-effort; a fault silently
     // suppresses the per-node sub-table rather than degrading the dashboard.
-    private async Task LoadSiteCallNodeKpis()
+    private async Task LoadSiteCallNodeKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetPerNodeSiteCallKpisAsync(
+            using var timeout = new CancellationTokenSource(TileTimeout);
-                new PerNodeSiteCallKpiRequest(Guid.NewGuid().ToString("N")));
+            var response = await KpiSnapshots.GetPerNodeSiteCallKpisAsync(forceRefresh, timeout.Token);
             if (response.Success)
             {
                 _siteCallNodeKpis = response.Nodes;
@@ -864,14 +893,20 @@
     // Audit KPI loader: wraps the service call so a transient DB outage degrades
     // the three tiles to em dashes with an inline error rather than killing the
     // dashboard. Mirrors LoadOutboxKpis's error handling shape.
-    private async Task LoadAuditKpis()
+    private async Task LoadAuditKpis(bool forceRefresh)
     {
         try
         {
-            _auditKpi = await AuditLogQueryService.GetKpiSnapshotAsync();
+            using var timeout = new CancellationTokenSource(TileTimeout);
+            _auditKpi = await KpiSnapshots.GetAuditKpiSnapshotAsync(forceRefresh, timeout.Token);
             _auditKpiAvailable = true;
             _auditKpiError = null;
         }
+        catch (OperationCanceledException)
+        {
+            _auditKpiAvailable = false;
+            _auditKpiError = "KPI query timed out.";
+        }
         catch (Exception ex)
         {
             _auditKpiAvailable = false;

src/ZB.MOM.WW.ScadaBridge.CentralUI/Components/Pages/Notifications/NotificationKpis.razor

@@ -7,7 +7,7 @@
 @using ZB.MOM.WW.ScadaBridge.Commons.Types.Kpi
 @using ZB.MOM.WW.ScadaBridge.Communication
 @using ZB.MOM.WW.ScadaBridge.CentralUI.Services
-@inject CommunicationService CommunicationService
+@inject IKpiSnapshotCache KpiSnapshots
 @inject ISiteRepository SiteRepository
 @inject IKpiHistoryQueryService KpiHistory
 @inject ILogger<NotificationKpis> Logger
@@ -250,15 +250,24 @@
             Logger.LogWarning(ex, "Failed to load sites for the KPI per-site breakdown.");
         }
 
-        await RefreshAll();
+        // First load rides whatever the shared KPI cache already holds fresh; the
+        // Refresh button forces a new query (see RefreshAll).
+        await LoadAllAsync(forceRefresh: false);
     }
 
-    private async Task RefreshAll()
+    // The operator's explicit Refresh bypasses the shared cache's freshness window.
+    private Task RefreshAll() => LoadAllAsync(forceRefresh: true);
+
+    private async Task LoadAllAsync(bool forceRefresh)
     {
         _loading = true;
         // Race-free despite all tasks mutating component fields: Blazor Server runs
         // every continuation on the circuit's single-threaded synchronization context.
-        await Task.WhenAll(LoadGlobalKpis(), LoadPerSiteKpis(), LoadPerNodeKpis(), LoadTrends());
+        await Task.WhenAll(
+            LoadGlobalKpis(forceRefresh),
+            LoadPerSiteKpis(forceRefresh),
+            LoadPerNodeKpis(forceRefresh),
+            LoadTrends());
         _loading = false;
     }
 
@@ -320,12 +329,11 @@
         }
     }
 
-    private async Task LoadGlobalKpis()
+    private async Task LoadGlobalKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetNotificationKpisAsync(
+            var response = await KpiSnapshots.GetNotificationKpisAsync(forceRefresh);
-                new NotificationKpiRequest(Guid.NewGuid().ToString("N")));
             if (response.Success)
             {
                 _kpi = response;
@@ -342,12 +350,11 @@
         }
     }
 
-    private async Task LoadPerSiteKpis()
+    private async Task LoadPerSiteKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetPerSiteNotificationKpisAsync(
+            var response = await KpiSnapshots.GetPerSiteNotificationKpisAsync(forceRefresh);
-                new PerSiteNotificationKpiRequest(Guid.NewGuid().ToString("N")));
             if (response.Success)
             {
                 _perSite = response.Sites;
@@ -364,12 +371,11 @@
         }
     }
 
-    private async Task LoadPerNodeKpis()
+    private async Task LoadPerNodeKpis(bool forceRefresh)
     {
         try
         {
-            var response = await CommunicationService.GetPerNodeNotificationKpisAsync(
+            var response = await KpiSnapshots.GetPerNodeNotificationKpisAsync(forceRefresh);
-                new PerNodeNotificationKpiRequest(Guid.NewGuid().ToString("N")));
             if (response.Success)
             {
                 _perNode = response.Nodes;

src/ZB.MOM.WW.ScadaBridge.CentralUI/ScriptAnalysis/ScriptAnalysisService.cs

@@ -179,6 +179,31 @@ public class ScriptAnalysisService
 
     private const int SandboxMaxCallSharedDepth = 16;
 
+    /// <summary>
+    /// Process-wide cap on concurrently executing Test Runs (arch-review WP2.4).
+    /// <para>
+    /// A sandbox run occupies a thread-pool thread for its whole duration — the
+    /// <c>SandboxScriptHost</c> attribute accessors are synchronous by contract and block on
+    /// cross-site I/O (<c>GetAwaiter().GetResult()</c>), so the thread is <em>parked</em>, not
+    /// merely busy, for up to the 10s sandbox timeout. Unbounded, a handful of Designers
+    /// each hitting Run on a slow-bound script could park enough pool threads to starve the
+    /// Blazor circuits sharing this process. The gate bounds that: excess runs wait for a
+    /// slot instead of taking a thread, and the wait is charged to the caller's own timeout.
+    /// </para>
+    /// <para>
+    /// Static because <c>ScriptAnalysisService</c> is registered scoped — a per-instance gate
+    /// would be per-request and bound nothing.
+    /// </para>
+    /// </summary>
+    private static readonly SemaphoreSlim SandboxRunGate =
+        new(SandboxMaxConcurrentRuns, SandboxMaxConcurrentRuns);
+
+    /// <summary>
+    /// Slot count for <see cref="SandboxRunGate"/>. Deliberately small: Test Run is an
+    /// authoring convenience on a node whose real job is serving circuits.
+    /// </summary>
+    private const int SandboxMaxConcurrentRuns = 4;
+
     /// <summary>
     /// Compiles and runs a script in the central process. The globals surface
     /// depends on <see cref="SandboxRunRequest.Kind"/>: template and shared
@@ -428,8 +453,15 @@ public class ScriptAnalysisService
         using var errorScope = captureError.BeginCapture(captured);
 
         var stopwatch = Stopwatch.StartNew();
+        var gateAcquired = false;
         try
         {
+            // Bound concurrent Test Runs (see SandboxRunGate). The wait is charged to the
+            // caller's own linked token, so queue time counts against the sandbox timeout
+            // and a queued run reports Timeout rather than hanging.
+            await SandboxRunGate.WaitAsync(linkedCts.Token).ConfigureAwait(false);
+            gateAcquired = true;
+
             // Run on a thread-pool thread with no SynchronizationContext: a
             // bound script's Instance.SetAttribute / Attributes[...] block
             // synchronously on cross-site I/O (the API surface is sync by
@@ -480,6 +512,10 @@ public class ScriptAnalysisService
                 $"{inner.GetType().Name}: {inner.Message}",
                 SandboxErrorKind.RuntimeError, stopwatch.ElapsedMilliseconds, null);
         }
+        finally
+        {
+            if (gateAcquired) SandboxRunGate.Release();
+        }
         // outScope / errorScope are disposed by their `using` declarations when the
         // method returns, restoring the previous capture scope on this call-tree
         // without touching process-global Console state.

src/ZB.MOM.WW.ScadaBridge.CentralUI/ServiceCollectionExtensions.cs

@@ -99,7 +99,17 @@ public static class ServiceCollectionExtensions
         // facade over CommunicationService.RequestDebugSnapshotAsync (the same
         // single-shot Ask the Debug View uses) — and flattens the alarm states.
         services.AddScoped<IInstanceSnapshotClient, CommunicationInstanceSnapshotClient>();
-        services.AddScoped<IAlarmSummaryService, AlarmSummaryService>();
+        // The fan-out itself stays scoped (it holds scoped repositories + the snapshot
+        // client), but the page resolves it through SharedAlarmSummaryService — a process
+        // singleton that memoizes the fan-out per site with single-flight, so N circuits
+        // watching one site cost ONE fan-out per window instead of N (arch-review WP2.4).
+        services.AddScoped<AlarmSummaryService>();
+        services.AddSingleton<IAlarmSummaryService, SharedAlarmSummaryService>();
+
+        // Process-level memoized point-in-time KPI snapshots (arch-review WP2.4, finding
+        // #8's per-circuit half). Singleton on purpose: the memo has to outlive any one
+        // circuit for N dashboards to share a single SQL round per refresh window.
+        services.AddSingleton<IKpiSnapshotCache, KpiSnapshotCache>();
 
         // Secured Writes: dispatches the two-person secured-write commands
         // (submit / approve / reject / list) to the central ManagementActor through the

src/ZB.MOM.WW.ScadaBridge.CentralUI/Services/AlarmSummaryService.cs

@@ -128,7 +128,20 @@ public sealed class AlarmSummaryService : IAlarmSummaryService
     }
 
     /// <inheritdoc/>
-    public AlarmSummaryResult BuildFromLiveAlarms(IReadOnlyList<AlarmStateChanged> alarms)
+    public AlarmSummaryResult BuildFromLiveAlarms(IReadOnlyList<AlarmStateChanged> alarms) =>
+        BuildFromLiveAlarmsCore(alarms);
+
+    /// <inheritdoc/>
+    public AlarmRollup ComputeRollup(IReadOnlyList<AlarmSummaryRow> rows) => ComputeRollupCore(rows);
+
+    /// <summary>
+    /// Pure flattening of a live-cache snapshot into summary rows. Shared with
+    /// <see cref="SharedAlarmSummaryService"/>, which memoizes the fan-out but must produce
+    /// byte-identical rows on the live path — one implementation, no drift.
+    /// </summary>
+    /// <param name="alarms">The live-cache alarm snapshot.</param>
+    /// <returns>The flattened, deterministically ordered rows (never any not-reporting names).</returns>
+    internal static AlarmSummaryResult BuildFromLiveAlarmsCore(IReadOnlyList<AlarmStateChanged> alarms)
     {
         ArgumentNullException.ThrowIfNull(alarms);
 
@@ -147,8 +160,14 @@ public sealed class AlarmSummaryService : IAlarmSummaryService
         return new AlarmSummaryResult(orderedRows, Array.Empty<string>());
     }
 
-    /// <inheritdoc/>
+    /// <summary>
-    public AlarmRollup ComputeRollup(IReadOnlyList<AlarmSummaryRow> rows)
+    /// Pure roll-up over already-flattened rows. Shared with
+    /// <see cref="SharedAlarmSummaryService"/> for the same one-implementation reason as
+    /// <see cref="BuildFromLiveAlarmsCore"/>.
+    /// </summary>
+    /// <param name="rows">The rows to roll up.</param>
+    /// <returns>The active/worst-severity/unacked/by-kind roll-up.</returns>
+    internal static AlarmRollup ComputeRollupCore(IReadOnlyList<AlarmSummaryRow> rows)
     {
         ArgumentNullException.ThrowIfNull(rows);
 

src/ZB.MOM.WW.ScadaBridge.CentralUI/Services/IKpiSnapshotCache.cs

@@ -0,0 +1,70 @@
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Audit;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Notification;
+using ZB.MOM.WW.ScadaBridge.Commons.Types;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+
+/// <summary>
+/// Process-level memoized point-in-time KPI snapshots shared by every Blazor circuit on
+/// this node (arch-review WP2.4, finding #8 — the per-circuit polling half).
+/// <para>
+/// Every KPI here is a <em>global, point-in-time aggregate</em>: it does not vary by user,
+/// site scope, or role, so N circuits asking for it inside one refresh window are asking
+/// the identical question. Without memoization, ten Health dashboards meant ten Asks →
+/// ten aggregate SQL round trips every ten seconds. Each accessor is independently
+/// memoized with a short TTL and single-flight production, so the fleet costs ONE round
+/// trip per KPI per window regardless of viewer count.
+/// </para>
+/// <para>
+/// Faults are never memoized: a failed query is re-attempted by the next caller, and the
+/// caller still sees the exception exactly as it would from a direct query — the pages'
+/// existing per-tile "unavailable" degradation is unchanged.
+/// </para>
+/// </summary>
+public interface IKpiSnapshotCache
+{
+    /// <summary>Global Notification Outbox KPIs (queue depth, stuck, parked, delivered-last-interval).</summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value (operator "Refresh"); an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized notification KPI response.</returns>
+    Task<NotificationKpiResponse> GetNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+
+    /// <summary>Per-source-site Notification Outbox KPI breakdown.</summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value; an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized per-site notification KPI response.</returns>
+    Task<PerSiteNotificationKpiResponse> GetPerSiteNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+
+    /// <summary>Per-node Notification Outbox KPI breakdown (grouped by <c>SourceNode</c>).</summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value; an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized per-node notification KPI response.</returns>
+    Task<PerNodeNotificationKpiResponse> GetPerNodeNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+
+    /// <summary>Global Site Call Audit KPIs (buffered, stuck, parked).</summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value; an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized site call KPI response.</returns>
+    Task<SiteCallKpiResponse> GetSiteCallKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+
+    /// <summary>Per-node Site Call Audit KPI breakdown (grouped by <c>SourceNode</c>).</summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value; an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized per-node site call KPI response.</returns>
+    Task<PerNodeSiteCallKpiResponse> GetPerNodeSiteCallKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+
+    /// <summary>
+    /// Audit Log KPI snapshot (1h volume + error rate over the central <c>AuditLog</c> table,
+    /// plus the summed per-site backlog).
+    /// </summary>
+    /// <param name="forceRefresh">Bypass a still-fresh value; an in-flight round is still joined.</param>
+    /// <param name="cancellationToken">Cancels this caller's wait, never the shared flight.</param>
+    /// <returns>The memoized audit KPI snapshot.</returns>
+    Task<AuditLogKpiSnapshot> GetAuditKpiSnapshotAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default);
+}

src/ZB.MOM.WW.ScadaBridge.CentralUI/Services/KpiSnapshotCache.cs

@@ -0,0 +1,122 @@
+using Microsoft.Extensions.DependencyInjection;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Audit;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Notification;
+using ZB.MOM.WW.ScadaBridge.Commons.Types;
+using ZB.MOM.WW.ScadaBridge.Communication;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+
+/// <summary>
+/// Default <see cref="IKpiSnapshotCache"/> — a DI <b>singleton</b> holding one
+/// <see cref="SingleFlightMemo{T}"/> per KPI query.
+/// <para>
+/// Registered as a singleton on purpose: the whole point is that the memo outlives any one
+/// circuit. It depends only on <see cref="CommunicationService"/> (itself a singleton) and
+/// <see cref="IServiceScopeFactory"/> — the audit KPI resolves its scoped query service from
+/// a fresh scope per flight, the same shape <c>AuditLogQueryService</c> uses internally so a
+/// KPI round never shares a circuit-scoped <c>DbContext</c>.
+/// </para>
+/// </summary>
+public sealed class KpiSnapshotCache : IKpiSnapshotCache
+{
+    /// <summary>
+    /// Freshness window. Sized just under the Health dashboard's 10s tick so each tick still
+    /// yields ONE fresh query for the whole node (rather than one per circuit), while a
+    /// circuit whose tick lands mid-window is served the memoized value instead of duplicating it.
+    /// </summary>
+    internal static readonly TimeSpan Ttl = TimeSpan.FromSeconds(8);
+
+    private readonly CommunicationService _communication;
+    private readonly IServiceScopeFactory _scopeFactory;
+
+    private readonly SingleFlightMemo<NotificationKpiResponse> _notificationKpis;
+    private readonly SingleFlightMemo<PerSiteNotificationKpiResponse> _perSiteNotificationKpis;
+    private readonly SingleFlightMemo<PerNodeNotificationKpiResponse> _perNodeNotificationKpis;
+    private readonly SingleFlightMemo<SiteCallKpiResponse> _siteCallKpis;
+    private readonly SingleFlightMemo<PerNodeSiteCallKpiResponse> _perNodeSiteCallKpis;
+    private readonly SingleFlightMemo<AuditLogKpiSnapshot> _auditKpis;
+
+    /// <summary>
+    /// Initializes the shared KPI snapshot cache.
+    /// </summary>
+    /// <param name="communication">Central-side comms used to Ask the outbox / site-call singletons.</param>
+    /// <param name="scopeFactory">Opens a fresh DI scope per audit-KPI flight.</param>
+    public KpiSnapshotCache(CommunicationService communication, IServiceScopeFactory scopeFactory)
+        : this(communication, scopeFactory, Ttl, clock: null)
+    {
+    }
+
+    /// <summary>
+    /// Test seam: same cache with an explicit TTL and clock so freshness can be asserted
+    /// without real delays.
+    /// </summary>
+    /// <param name="communication">Central-side comms used to Ask the outbox / site-call singletons.</param>
+    /// <param name="scopeFactory">Opens a fresh DI scope per audit-KPI flight.</param>
+    /// <param name="ttl">Freshness window.</param>
+    /// <param name="clock">Clock used for freshness.</param>
+    internal KpiSnapshotCache(
+        CommunicationService communication,
+        IServiceScopeFactory scopeFactory,
+        TimeSpan ttl,
+        Func<DateTimeOffset>? clock)
+    {
+        _communication = communication ?? throw new ArgumentNullException(nameof(communication));
+        _scopeFactory = scopeFactory ?? throw new ArgumentNullException(nameof(scopeFactory));
+
+        _notificationKpis = new SingleFlightMemo<NotificationKpiResponse>(ttl, clock);
+        _perSiteNotificationKpis = new SingleFlightMemo<PerSiteNotificationKpiResponse>(ttl, clock);
+        _perNodeNotificationKpis = new SingleFlightMemo<PerNodeNotificationKpiResponse>(ttl, clock);
+        _siteCallKpis = new SingleFlightMemo<SiteCallKpiResponse>(ttl, clock);
+        _perNodeSiteCallKpis = new SingleFlightMemo<PerNodeSiteCallKpiResponse>(ttl, clock);
+        _auditKpis = new SingleFlightMemo<AuditLogKpiSnapshot>(ttl, clock);
+    }
+
+    /// <inheritdoc/>
+    public Task<NotificationKpiResponse> GetNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _notificationKpis.GetAsync(
+            () => _communication.GetNotificationKpisAsync(new NotificationKpiRequest(NewCorrelationId())),
+            forceRefresh, cancellationToken);
+
+    /// <inheritdoc/>
+    public Task<PerSiteNotificationKpiResponse> GetPerSiteNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _perSiteNotificationKpis.GetAsync(
+            () => _communication.GetPerSiteNotificationKpisAsync(new PerSiteNotificationKpiRequest(NewCorrelationId())),
+            forceRefresh, cancellationToken);
+
+    /// <inheritdoc/>
+    public Task<PerNodeNotificationKpiResponse> GetPerNodeNotificationKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _perNodeNotificationKpis.GetAsync(
+            () => _communication.GetPerNodeNotificationKpisAsync(new PerNodeNotificationKpiRequest(NewCorrelationId())),
+            forceRefresh, cancellationToken);
+
+    /// <inheritdoc/>
+    public Task<SiteCallKpiResponse> GetSiteCallKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _siteCallKpis.GetAsync(
+            () => _communication.GetSiteCallKpisAsync(new SiteCallKpiRequest(NewCorrelationId())),
+            forceRefresh, cancellationToken);
+
+    /// <inheritdoc/>
+    public Task<PerNodeSiteCallKpiResponse> GetPerNodeSiteCallKpisAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _perNodeSiteCallKpis.GetAsync(
+            () => _communication.GetPerNodeSiteCallKpisAsync(new PerNodeSiteCallKpiRequest(NewCorrelationId())),
+            forceRefresh, cancellationToken);
+
+    /// <inheritdoc/>
+    public Task<AuditLogKpiSnapshot> GetAuditKpiSnapshotAsync(
+        bool forceRefresh = false, CancellationToken cancellationToken = default) =>
+        _auditKpis.GetAsync(LoadAuditKpisAsync, forceRefresh, cancellationToken);
+
+    private async Task<AuditLogKpiSnapshot> LoadAuditKpisAsync()
+    {
+        await using var scope = _scopeFactory.CreateAsyncScope();
+        var queryService = scope.ServiceProvider.GetRequiredService<IAuditLogQueryService>();
+        return await queryService.GetKpiSnapshotAsync();
+    }
+
+    private static string NewCorrelationId() => Guid.NewGuid().ToString("N");
+}

src/ZB.MOM.WW.ScadaBridge.CentralUI/Services/SharedAlarmSummaryService.cs

@@ -0,0 +1,115 @@
+using System.Collections.Concurrent;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Options;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Streaming;
+using ZB.MOM.WW.ScadaBridge.Communication;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+
+/// <summary>
+/// Process-level memoizing façade over <see cref="AlarmSummaryService"/> (arch-review WP2.4).
+/// <para>
+/// The Alarm Summary page polls every 15s <em>per circuit</em>, and each poll fans one debug
+/// snapshot Ask out to every Enabled instance on the site. Ten operators watching one site
+/// meant ten independent fan-outs per 15s across the same instances — on top of the shared
+/// live-alarm aggregator already seeding and reconciling from the identical fan-out. The
+/// question is not user-specific, so this service collapses it: one memo slot per site,
+/// single-flight, so the node costs ONE fan-out per site per window no matter how many
+/// viewers are watching.
+/// </para>
+/// <para>
+/// <b>The freshness window follows the live cache.</b> While
+/// <see cref="ISiteAlarmLiveCache.IsLive"/> is true the page deliberately ignores the poll's
+/// alarm rows (the live deltas own them — arch-review R2 N5), so the only thing the poll
+/// still supplies is the not-reporting list, and the window widens to the aggregator's own
+/// reconcile interval. While the cache is cold the poll is the page's full-rebuild safety
+/// net, so the window stays just under the page's 15s tick and every tick gets fresh data.
+/// </para>
+/// </summary>
+public sealed class SharedAlarmSummaryService : IAlarmSummaryService
+{
+    /// <summary>
+    /// Freshness window while the live alarm cache is NOT serving this site: the poll is the
+    /// page's authoritative rebuild path, so it must stay under the 15s page tick.
+    /// </summary>
+    internal static readonly TimeSpan ColdCacheTtl = TimeSpan.FromSeconds(12);
+
+    private readonly IServiceScopeFactory _scopeFactory;
+    private readonly ISiteAlarmLiveCache _liveCache;
+    private readonly TimeSpan _liveCacheTtl;
+    private readonly Func<DateTimeOffset>? _clock;
+
+    private readonly ConcurrentDictionary<int, SingleFlightMemo<AlarmSummaryResult>> _bySite = new();
+
+    /// <summary>
+    /// Initializes the shared alarm summary façade.
+    /// </summary>
+    /// <param name="scopeFactory">Opens a fresh DI scope per fan-out (fresh repositories, off any circuit scope).</param>
+    /// <param name="liveCache">The shared live alarm cache, consulted only for its per-site liveness.</param>
+    /// <param name="options">Communication options; supplies the aggregator reconcile interval.</param>
+    public SharedAlarmSummaryService(
+        IServiceScopeFactory scopeFactory,
+        ISiteAlarmLiveCache liveCache,
+        IOptions<CommunicationOptions> options)
+        : this(scopeFactory, liveCache, (options ?? throw new ArgumentNullException(nameof(options)))
+            .Value.LiveAlarmCacheReconcileInterval, clock: null)
+    {
+    }
+
+    /// <summary>
+    /// Test seam: same façade with an explicit live-cache window and clock.
+    /// </summary>
+    /// <param name="scopeFactory">Opens a fresh DI scope per fan-out.</param>
+    /// <param name="liveCache">The shared live alarm cache.</param>
+    /// <param name="liveCacheTtl">Freshness window used while the live cache is serving the site.</param>
+    /// <param name="clock">Clock used for freshness.</param>
+    internal SharedAlarmSummaryService(
+        IServiceScopeFactory scopeFactory,
+        ISiteAlarmLiveCache liveCache,
+        TimeSpan liveCacheTtl,
+        Func<DateTimeOffset>? clock)
+    {
+        _scopeFactory = scopeFactory ?? throw new ArgumentNullException(nameof(scopeFactory));
+        _liveCache = liveCache ?? throw new ArgumentNullException(nameof(liveCache));
+        // Never shorter than the cold window — a misconfigured reconcile interval must not
+        // silently make the memo useless.
+        _liveCacheTtl = liveCacheTtl > ColdCacheTtl ? liveCacheTtl : ColdCacheTtl;
+        _clock = clock;
+    }
+
+    /// <inheritdoc/>
+    public Task<AlarmSummaryResult> GetSiteAlarmsAsync(
+        int siteId, CancellationToken cancellationToken = default)
+    {
+        var memo = _bySite.GetOrAdd(
+            siteId,
+            _ => new SingleFlightMemo<AlarmSummaryResult>(ColdCacheTtl, _clock));
+
+        var ttl = _liveCache.IsLive(siteId) ? _liveCacheTtl : ColdCacheTtl;
+
+        return memo.GetAsync(
+            () => FanOutAsync(siteId),
+            forceRefresh: false,
+            cancellationToken,
+            ttlOverride: ttl);
+    }
+
+    /// <inheritdoc/>
+    public AlarmSummaryResult BuildFromLiveAlarms(IReadOnlyList<AlarmStateChanged> alarms) =>
+        AlarmSummaryService.BuildFromLiveAlarmsCore(alarms);
+
+    /// <inheritdoc/>
+    public AlarmRollup ComputeRollup(IReadOnlyList<AlarmSummaryRow> rows) =>
+        AlarmSummaryService.ComputeRollupCore(rows);
+
+    /// <summary>
+    /// One shared fan-out. Deliberately runs with no caller cancellation token: the flight is
+    /// shared, so one circuit navigating away must not cancel the round the others are awaiting.
+    /// </summary>
+    private async Task<AlarmSummaryResult> FanOutAsync(int siteId)
+    {
+        await using var scope = _scopeFactory.CreateAsyncScope();
+        var inner = scope.ServiceProvider.GetRequiredService<AlarmSummaryService>();
+        return await inner.GetSiteAlarmsAsync(siteId, CancellationToken.None);
+    }
+}

src/ZB.MOM.WW.ScadaBridge.CentralUI/Services/SingleFlightMemo.cs

@@ -0,0 +1,136 @@
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+
+/// <summary>
+/// A process-level memo slot: one value, a time-to-live, and single-flight production.
+/// <para>
+/// Blazor Server runs one circuit per connected browser, so a page that polls on a timer
+/// runs its query <em>per circuit</em>. Ten operators on the Health dashboard meant ten
+/// identical KPI round trips every ten seconds. This slot collapses them: the first caller
+/// in a TTL window starts the flight, every concurrent caller awaits the <em>same</em>
+/// <see cref="Task{TResult}"/>, and callers arriving after it completes are served the
+/// memoized value until the TTL expires.
+/// </para>
+/// <para>
+/// Failures are never cached — a faulted flight is discarded so the next caller retries.
+/// The caller's own exception handling is therefore unchanged: it still sees the fault
+/// (all callers of that one flight see it), and the next window starts clean.
+/// </para>
+/// </summary>
+/// <typeparam name="T">The memoized value type.</typeparam>
+internal sealed class SingleFlightMemo<T>
+{
+    private readonly object _lock = new();
+    private readonly TimeSpan _ttl;
+    private readonly Func<DateTimeOffset> _clock;
+
+    /// <summary>The current flight, completed or not. Null until the first call.</summary>
+    private Task<T>? _current;
+
+    /// <summary>Wall-clock instant after which <see cref="_current"/> is stale. Only meaningful when it completed successfully.</summary>
+    private DateTimeOffset _freshUntil;
+
+    /// <summary>
+    /// Initializes a memo slot.
+    /// </summary>
+    /// <param name="ttl">How long a successfully produced value stays fresh.</param>
+    /// <param name="clock">Clock used for freshness, injectable so tests need no real delay.</param>
+    public SingleFlightMemo(TimeSpan ttl, Func<DateTimeOffset>? clock = null)
+    {
+        _ttl = ttl;
+        _clock = clock ?? (() => DateTimeOffset.UtcNow);
+    }
+
+    /// <summary>
+    /// Number of times the factory has actually been invoked. Diagnostics and tests only.
+    /// </summary>
+    public int FlightCount { get; private set; }
+
+    /// <summary>
+    /// Returns the memoized value, starting a flight if none is fresh.
+    /// </summary>
+    /// <param name="factory">Produces a fresh value. Invoked outside the slot's lock.</param>
+    /// <param name="forceRefresh">
+    /// When <see langword="true"/>, a completed-but-still-fresh value is discarded and a new
+    /// flight starts. An <em>in-flight</em> round is still joined rather than duplicated, so a
+    /// burst of operator "Refresh" clicks remains one round trip.
+    /// </param>
+    /// <param name="cancellationToken">
+    /// Cancels this caller's wait only — never the shared flight, which other callers may still
+    /// be awaiting.
+    /// </param>
+    /// <param name="ttlOverride">
+    /// Freshness window for <em>this</em> flight, overriding the slot's default. Used where the
+    /// acceptable staleness depends on state the caller knows (the Alarm Summary poll needs a
+    /// far shorter window while the live alarm cache is cold than while it is serving deltas).
+    /// </param>
+    /// <returns>The memoized (or freshly produced) value.</returns>
+    public Task<T> GetAsync(
+        Func<Task<T>> factory,
+        bool forceRefresh = false,
+        CancellationToken cancellationToken = default,
+        TimeSpan? ttlOverride = null)
+    {
+        ArgumentNullException.ThrowIfNull(factory);
+
+        TaskCompletionSource<T>? started = null;
+        Task<T> result;
+
+        lock (_lock)
+        {
+            var current = _current;
+            if (current is { IsCompleted: false })
+            {
+                // A round is already in progress — join it, force or not.
+                result = current;
+            }
+            else if (!forceRefresh
+                && current is { IsCompletedSuccessfully: true }
+                && _clock() < _freshUntil)
+            {
+                result = current;
+            }
+            else
+            {
+                // RunContinuationsAsynchronously: a caller's continuation must never run
+                // inline on the thread that completes the flight (that thread is another
+                // circuit's dispatcher).
+                started = new TaskCompletionSource<T>(TaskCreationOptions.RunContinuationsAsynchronously);
+                _current = started.Task;
+                FlightCount++;
+                result = started.Task;
+            }
+        }
+
+        // The factory runs OUTSIDE the lock: it does I/O (an Ask, a SQL round trip) and must
+        // never be able to re-enter this slot on the calling thread while the lock is held.
+        if (started is not null)
+        {
+            _ = RunAsync(factory, started, ttlOverride ?? _ttl);
+        }
+
+        return cancellationToken.CanBeCanceled ? result.WaitAsync(cancellationToken) : result;
+    }
+
+    private async Task RunAsync(Func<Task<T>> factory, TaskCompletionSource<T> completion, TimeSpan ttl)
+    {
+        try
+        {
+            var value = await factory().ConfigureAwait(false);
+            lock (_lock)
+            {
+                _freshUntil = _clock() + ttl;
+            }
+            completion.SetResult(value);
+        }
+        catch (Exception ex)
+        {
+            // A failure is not memoized: leave _freshUntil in the past so the next caller
+            // starts a new flight instead of re-serving the fault for the whole TTL.
+            lock (_lock)
+            {
+                _freshUntil = default;
+            }
+            completion.SetException(ex);
+        }
+    }
+}

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Deployment/DebugViewRenderCoalescingTests.cs

@@ -0,0 +1,144 @@
+using System.Collections;
+using System.Reflection;
+using System.Security.Claims;
+using Bunit;
+using Microsoft.AspNetCore.Components.Authorization;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using NSubstitute;
+using ZB.MOM.WW.ScadaBridge.CentralUI.Auth;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Sites;
+using ZB.MOM.WW.ScadaBridge.Commons.Interfaces.Repositories;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Streaming;
+using ZB.MOM.WW.ScadaBridge.Communication;
+using ZB.MOM.WW.ScadaBridge.Communication.Grpc;
+using DebugViewPage = ZB.MOM.WW.ScadaBridge.CentralUI.Components.Pages.Deployment.DebugView;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Tests.Deployment;
+
+/// <summary>
+/// Regression tests for the Debug View render coalescing (arch-review WP2.4). Every streamed
+/// event used to marshal onto the circuit dispatcher and call <c>StateHasChanged</c> on its
+/// own, so a chatty instance drove one full render — and two full composition-tree rebuilds —
+/// per value change. Events are now buffered and applied once per ~250 ms window: a burst
+/// costs a handful of renders instead of one per event, and no event is dropped.
+/// </summary>
+public class DebugViewRenderCoalescingTests : BunitContext
+{
+    private IRenderedComponent<DebugViewPage> RenderPage()
+    {
+        JSInterop.Mode = JSRuntimeMode.Loose;
+
+        var repo = Substitute.For<ITemplateEngineRepository>();
+        var siteRepo = Substitute.For<ISiteRepository>();
+        siteRepo.GetAllSitesAsync().Returns(new List<Site>());
+        Services.AddSingleton(repo);
+        Services.AddSingleton(siteRepo);
+
+        var comms = new CommunicationService(
+            Options.Create(new CommunicationOptions()),
+            NullLogger<CommunicationService>.Instance);
+        Services.AddSingleton(comms);
+
+        var grpcFactory = new SiteStreamGrpcClientFactory(NullLoggerFactory.Instance);
+        var debugStream = new DebugStreamService(
+            comms, new ServiceCollection().BuildServiceProvider(), grpcFactory,
+            NullLogger<DebugStreamService>.Instance);
+        Services.AddSingleton(debugStream);
+
+        var identity = new ClaimsIdentity(
+            new[] { new Claim(ClaimTypes.Name, "deployer") }, "TestCookie");
+        var stubAuth = new StubAuthStateProvider(
+            new AuthenticationState(new ClaimsPrincipal(identity)));
+        Services.AddSingleton<AuthenticationStateProvider>(stubAuth);
+        Services.AddScoped(_ => new SiteScopeService(stubAuth));
+
+        return Render<DebugViewPage>();
+    }
+
+    private sealed class StubAuthStateProvider : AuthenticationStateProvider
+    {
+        private readonly AuthenticationState _state;
+        public StubAuthStateProvider(AuthenticationState state) => _state = state;
+        public override Task<AuthenticationState> GetAuthenticationStateAsync()
+            => Task.FromResult(_state);
+    }
+
+    private static MethodInfo HandleStreamEvent => typeof(DebugViewPage).GetMethod(
+        "HandleStreamEvent", BindingFlags.Instance | BindingFlags.NonPublic)!;
+
+    private static IDictionary AttributeValues(DebugViewPage c) => (IDictionary)
+        typeof(DebugViewPage).GetField("_attributeValues",
+            BindingFlags.Instance | BindingFlags.NonPublic)!.GetValue(c)!;
+
+    private static void Send(DebugViewPage page, string name, object value) =>
+        HandleStreamEvent.Invoke(page, new object[]
+        {
+            new AttributeValueChanged("Inst-1", name, name, value, "Good", DateTimeOffset.UtcNow),
+        });
+
+    [Fact]
+    public void BurstOfEvents_RendersOnce_ButAppliesEveryEvent()
+    {
+        var cut = RenderPage();
+        var dict = AttributeValues(cut.Instance);
+        var rendersBefore = cut.RenderCount;
+
+        const int burst = 300;
+        for (var i = 0; i < burst; i++)
+        {
+            Send(cut.Instance, $"Tag.{i}", i);
+        }
+
+        cut.WaitForState(() => dict.Count == burst, TimeSpan.FromSeconds(5));
+
+        // Pre-fix this was one render per event. The window may close mid-burst, so allow a
+        // few flushes — the point is that it is a small constant, not O(events).
+        var renders = cut.RenderCount - rendersBefore;
+        Assert.InRange(renders, 1, 5);
+    }
+
+    [Fact]
+    public void RepeatedUpdatesToOneTag_CollapseToTheLatestValue()
+    {
+        var cut = RenderPage();
+        var dict = AttributeValues(cut.Instance);
+
+        for (var i = 0; i < 50; i++)
+        {
+            Send(cut.Instance, "Pump.Speed", i);
+        }
+
+        cut.WaitForState(() => dict.Count == 1, TimeSpan.FromSeconds(5));
+        var applied = (AttributeValueChanged)dict["Pump.Speed"]!;
+        Assert.Equal(49, applied.Value);
+    }
+
+    [Fact]
+    public void EventArrivingAfterAFlush_StillArmsANewWindow()
+    {
+        var cut = RenderPage();
+        var dict = AttributeValues(cut.Instance);
+
+        Send(cut.Instance, "First", 1);
+        cut.WaitForState(() => dict.Count == 1, TimeSpan.FromSeconds(5));
+
+        // The window disarms on drain; a later event must arm a fresh one rather than
+        // sitting in the pending map until some unrelated event happens to arrive.
+        Send(cut.Instance, "Second", 2);
+        cut.WaitForState(() => dict.Count == 2, TimeSpan.FromSeconds(5));
+    }
+
+    [Fact]
+    public void EventsAfterDispose_AreDroppedWithoutThrowing()
+    {
+        var cut = RenderPage();
+        cut.Instance.Dispose();
+
+        var ex = Record.Exception(() => Send(cut.Instance, "Late", 1));
+
+        Assert.Null(ex);
+        Assert.Empty(AttributeValues(cut.Instance));
+    }
+}

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Deployment/DeploymentsReloadDebounceTests.cs

@@ -0,0 +1,127 @@
+using System.Security.Claims;
+using Bunit;
+using Microsoft.AspNetCore.Components.Authorization;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using NSubstitute;
+using ZB.MOM.WW.ScadaBridge.CentralUI.Auth;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Deployment;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Instances;
+using ZB.MOM.WW.ScadaBridge.Commons.Interfaces.Repositories;
+using ZB.MOM.WW.ScadaBridge.Commons.Types.Enums;
+using ZB.MOM.WW.ScadaBridge.DeploymentManager;
+using DeploymentsPage = ZB.MOM.WW.ScadaBridge.CentralUI.Components.Pages.Deployment.Deployments;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Tests.Deployment;
+
+/// <summary>
+/// Regression tests for the Deployment Status push coalescing (arch-review WP2.4). One
+/// notifier callback reloads EVERY deployment record plus EVERY instance, and the notifier
+/// fires per status write — so a multi-instance deploy produced a stampede of full reloads
+/// per circuit. The reload is now leading-edge debounced: the first push after an idle gap
+/// is still immediate, and a burst behind it collapses into one trailing reload.
+/// </summary>
+public class DeploymentsReloadDebounceTests : BunitContext
+{
+    private IDeploymentManagerRepository _deployRepo = null!;
+    private ITemplateEngineRepository _templateRepo = null!;
+    private DeploymentStatusNotifier _notifier = null!;
+
+    private void RegisterServices()
+    {
+        _deployRepo = Substitute.For<IDeploymentManagerRepository>();
+        _templateRepo = Substitute.For<ITemplateEngineRepository>();
+        _notifier = new DeploymentStatusNotifier(NullLogger<DeploymentStatusNotifier>.Instance);
+
+        _templateRepo.GetAllInstancesAsync(Arg.Any<CancellationToken>())
+            .Returns(new List<Instance> { new("Inst-1") { Id = 1, SiteId = 1 } });
+        _deployRepo.GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>())
+            .Returns(new List<DeploymentRecord>());
+
+        Services.AddSingleton(_deployRepo);
+        Services.AddSingleton(_templateRepo);
+        Services.AddSingleton<IDeploymentStatusNotifier>(_notifier);
+
+        var identity = new ClaimsIdentity(
+            new[] { new Claim(ClaimTypes.Name, "deployer") }, "TestCookie");
+        var stubAuth = new StubAuthStateProvider(
+            new AuthenticationState(new ClaimsPrincipal(identity)));
+        Services.AddSingleton<AuthenticationStateProvider>(stubAuth);
+        Services.AddScoped(_ => new SiteScopeService(stubAuth));
+    }
+
+    private sealed class StubAuthStateProvider : AuthenticationStateProvider
+    {
+        private readonly AuthenticationState _state;
+        public StubAuthStateProvider(AuthenticationState state) => _state = state;
+        public override Task<AuthenticationState> GetAuthenticationStateAsync()
+            => Task.FromResult(_state);
+    }
+
+    [Fact]
+    public void BurstOfStatusWrites_CollapsesIntoFarFewerReloads()
+    {
+        RegisterServices();
+        var cut = Render<DeploymentsPage>();
+        cut.WaitForAssertion(() =>
+            _deployRepo.Received().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>()));
+        _deployRepo.ClearReceivedCalls();
+
+        // A 40-instance site deploy: every status write raises the notifier.
+        for (var i = 0; i < 40; i++)
+        {
+            _notifier.NotifyStatusChanged(
+                new DeploymentStatusChange($"dep-{i}", 1, DeploymentStatus.InProgress));
+        }
+
+        // Leading edge fires at once; the rest ride one trailing reload.
+        cut.WaitForAssertion(() =>
+            _deployRepo.Received().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>()));
+
+        // Let the trailing window close before counting.
+        Thread.Sleep(900);
+
+        var reloads = _deployRepo.ReceivedCalls()
+            .Count(c => c.GetMethodInfo().Name == nameof(IDeploymentManagerRepository.GetAllDeploymentRecordsAsync));
+        Assert.InRange(reloads, 1, 4);
+    }
+
+    [Fact]
+    public void FirstPushAfterIdle_ReloadsImmediately()
+    {
+        RegisterServices();
+        var cut = Render<DeploymentsPage>();
+        cut.WaitForAssertion(() =>
+            _deployRepo.Received().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>()));
+        _deployRepo.ClearReceivedCalls();
+
+        // Idle since the initial load — the leading edge must not wait out the window.
+        Thread.Sleep(600);
+        _notifier.NotifyStatusChanged(
+            new DeploymentStatusChange("dep-1", 1, DeploymentStatus.Success));
+
+        cut.WaitForAssertion(
+            () => _deployRepo.Received().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>()),
+            TimeSpan.FromMilliseconds(400));
+    }
+
+    [Fact]
+    public void DisposeDuringACoalesceWindow_DoesNotReload()
+    {
+        RegisterServices();
+        var cut = Render<DeploymentsPage>();
+        cut.WaitForAssertion(() =>
+            _deployRepo.Received().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>()));
+
+        // Two pushes: the first takes the leading edge, the second arms the trailing timer.
+        _notifier.NotifyStatusChanged(new DeploymentStatusChange("dep-1", 1, DeploymentStatus.InProgress));
+        _notifier.NotifyStatusChanged(new DeploymentStatusChange("dep-2", 1, DeploymentStatus.InProgress));
+
+        cut.Instance.Dispose();
+        _deployRepo.ClearReceivedCalls();
+
+        // The armed timer must be disposed with the component, not fire against it.
+        Thread.Sleep(900);
+        _deployRepo.DidNotReceive().GetAllDeploymentRecordsAsync(Arg.Any<CancellationToken>());
+    }
+}

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Monitoring/AlarmSummaryVirtualizeTests.cs

@@ -0,0 +1,113 @@
+using System.Security.Claims;
+using Bunit;
+using Microsoft.AspNetCore.Components.Authorization;
+using Microsoft.Extensions.DependencyInjection;
+using NSubstitute;
+using ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Sites;
+using ZB.MOM.WW.ScadaBridge.Commons.Interfaces.Repositories;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Streaming;
+using ZB.MOM.WW.ScadaBridge.Commons.Types.Enums;
+using ZB.MOM.WW.ScadaBridge.Communication;
+using ZB.MOM.WW.ScadaBridge.Security;
+using AlarmSummaryPage = ZB.MOM.WW.ScadaBridge.CentralUI.Components.Pages.Monitoring.AlarmSummary;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Tests.Monitoring;
+
+/// <summary>
+/// Covers the Alarm Summary flat table's two rendering paths (arch-review WP2.4). A short
+/// list renders every row with a plain foreach; past the threshold the same row template is
+/// handed to <c>Virtualize</c> so a site with thousands of alarms stops materialising the
+/// whole table into the circuit's render tree on every delta. Row markup — including the
+/// <c>data-test</c> hook the operator tests key off — is identical either way.
+/// </summary>
+public class AlarmSummaryVirtualizeTests : BunitContext
+{
+    private readonly IAlarmSummaryService _summary = Substitute.For<IAlarmSummaryService>();
+    private readonly ISiteRepository _siteRepo = Substitute.For<ISiteRepository>();
+
+    private void Arrange(int rowCount)
+    {
+        JSInterop.Mode = JSRuntimeMode.Loose;
+
+        var rows = Enumerable.Range(0, rowCount)
+            .Select(i => new AlarmSummaryRow(
+                $"inst-{i:D4}",
+                new AlarmStateChanged($"inst-{i:D4}", $"alarm-{i:D4}", AlarmState.Active, i, DateTimeOffset.UtcNow)))
+            .ToList();
+
+        _summary.GetSiteAlarmsAsync(Arg.Any<int>(), Arg.Any<CancellationToken>())
+            .Returns(Task.FromResult(new AlarmSummaryResult(rows, Array.Empty<string>())));
+        _summary.ComputeRollup(Arg.Any<IReadOnlyList<AlarmSummaryRow>>())
+            .Returns(new AlarmRollup(rowCount, rowCount, 0, new Dictionary<AlarmKind, int>()));
+        Services.AddSingleton(_summary);
+
+        _siteRepo.GetAllSitesAsync(Arg.Any<CancellationToken>())
+            .Returns(Task.FromResult<IReadOnlyList<Site>>(new List<Site>
+            {
+                new("Site 1", "site1") { Id = 1 },
+            }));
+        Services.AddSingleton(_siteRepo);
+        Services.AddSingleton<ISiteAlarmLiveCache>(new InertLiveCache());
+
+        var claims = new[]
+        {
+            new Claim(JwtTokenService.UsernameClaimType, "tester"),
+            new Claim(JwtTokenService.RoleClaimType, "Administrator"),
+        };
+        Services.AddSingleton<AuthenticationStateProvider>(
+            new TestAuthStateProvider(new ClaimsPrincipal(new ClaimsIdentity(claims, "TestAuth"))));
+        Services.AddAuthorizationCore();
+    }
+
+    private IRenderedComponent<AlarmSummaryPage> RenderWithSiteSelected()
+    {
+        var cut = Render<AlarmSummaryPage>();
+        cut.Find("[data-test='alarm-summary-site']").Change("1");
+        return cut;
+    }
+
+    [Fact]
+    public void ShortList_RendersEveryRowInline()
+    {
+        Arrange(rowCount: 20);
+
+        var cut = RenderWithSiteSelected();
+
+        cut.WaitForAssertion(() =>
+            Assert.Equal(20, cut.FindAll("tr[data-test='alarm-summary-row']").Count));
+        // The row-count line is unchanged by either path.
+        Assert.Contains("Showing 20 of 20", cut.Markup);
+    }
+
+    [Fact]
+    public void LongList_VirtualizesWithoutRenderingEveryRow()
+    {
+        Arrange(rowCount: 2000);
+
+        var cut = RenderWithSiteSelected();
+
+        cut.WaitForAssertion(() => Assert.Contains("Showing 2000 of 2000", cut.Markup));
+
+        var rendered = cut.FindAll("tr[data-test='alarm-summary-row']").Count;
+        Assert.InRange(rendered, 1, 1999);
+        // Rows still carry the same shape — instance name, alarm name, severity.
+        Assert.Contains("alarm-", cut.Markup);
+    }
+
+    /// <summary>Never goes live, so the page keeps its poll snapshot for these tests.</summary>
+    private sealed class InertLiveCache : ISiteAlarmLiveCache
+    {
+        public IDisposable Subscribe(int siteId, Action onChanged) => new NoOp();
+
+        public IReadOnlyList<AlarmStateChanged> GetCurrentAlarms(int siteId) =>
+            Array.Empty<AlarmStateChanged>();
+
+        public bool IsLive(int siteId) => false;
+
+        private sealed class NoOp : IDisposable
+        {
+            public void Dispose() { }
+        }
+    }
+}

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Pages/HealthPageTests.cs

@@ -64,6 +64,14 @@ public class HealthPageTests : BunitContext
         _comms.SetSiteCallAudit(siteCallAudit);
         Services.AddSingleton(_comms);
 
+        // arch-review WP2.4 — the page no longer queries CommunicationService /
+        // IAuditLogQueryService per circuit; it reads the process-level memoized
+        // KPI cache. The real cache is registered here (not a substitute) so the
+        // scripted-actor seam above is still what actually answers.
+        Services.AddSingleton<IKpiSnapshotCache>(sp => new KpiSnapshotCache(
+            sp.GetRequiredService<CommunicationService>(),
+            sp.GetRequiredService<IServiceScopeFactory>()));
+
         var aggregator = Substitute.For<ICentralHealthAggregator>();
         aggregator.GetAllSiteStates()
             .Returns(new Dictionary<string, SiteHealthState>());

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Pages/NotificationKpisPageTests.cs

@@ -71,6 +71,13 @@ public class NotificationKpisPageTests : BunitContext
 
         Services.AddSingleton(_comms);
 
+        // arch-review WP2.4 — the page reads the process-level memoized KPI cache
+        // rather than querying CommunicationService per circuit. The real cache is
+        // registered so the scripted-actor seam above still answers the queries.
+        Services.AddSingleton<IKpiSnapshotCache>(sp => new KpiSnapshotCache(
+            sp.GetRequiredService<CommunicationService>(),
+            sp.GetRequiredService<IServiceScopeFactory>()));
+
         var siteRepo = Substitute.For<ISiteRepository>();
         siteRepo.GetAllSitesAsync(Arg.Any<CancellationToken>())
             .Returns(Task.FromResult<IReadOnlyList<Site>>(new List<Site>

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Services/SharedAlarmSummaryServiceTests.cs

@@ -0,0 +1,169 @@
+using Microsoft.Extensions.DependencyInjection;
+using NSubstitute;
+using ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Instances;
+using ZB.MOM.WW.ScadaBridge.Commons.Entities.Sites;
+using ZB.MOM.WW.ScadaBridge.Commons.Interfaces.Repositories;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.DebugView;
+using ZB.MOM.WW.ScadaBridge.Commons.Messages.Streaming;
+using ZB.MOM.WW.ScadaBridge.Commons.Types.Enums;
+using ZB.MOM.WW.ScadaBridge.Communication;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Tests.Services;
+
+/// <summary>
+/// Unit tests for <see cref="SharedAlarmSummaryService"/> (arch-review WP2.4). The Alarm
+/// Summary page polls per circuit; the shared façade must collapse those into ONE per-site
+/// fan-out per freshness window, and must widen that window while the live alarm cache is
+/// serving the site (where the poll only still supplies the not-reporting list).
+/// </summary>
+public class SharedAlarmSummaryServiceTests : IDisposable
+{
+    private const int SiteId = 7;
+    private const string SiteIdentifier = "plant-a";
+
+    private static readonly DateTimeOffset T0 = new(2026, 8, 14, 9, 0, 0, TimeSpan.Zero);
+
+    private readonly ITemplateEngineRepository _instanceRepo = Substitute.For<ITemplateEngineRepository>();
+    private readonly ISiteRepository _siteRepo = Substitute.For<ISiteRepository>();
+    private readonly IInstanceSnapshotClient _snapshotClient = Substitute.For<IInstanceSnapshotClient>();
+    private readonly FakeLiveCache _liveCache = new();
+    private readonly ServiceProvider _provider;
+
+    private DateTimeOffset _now = T0;
+
+    public SharedAlarmSummaryServiceTests()
+    {
+        _siteRepo.GetSiteByIdAsync(SiteId, Arg.Any<CancellationToken>())
+            .Returns(new Site("Plant A", SiteIdentifier) { Id = SiteId });
+        _instanceRepo.GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>())
+            .Returns(new List<Instance>
+            {
+                new("inst-a") { Id = 1, SiteId = SiteId, State = InstanceState.Enabled },
+            });
+        _snapshotClient.GetSnapshotAsync(SiteIdentifier, "inst-a", Arg.Any<CancellationToken>())
+            .Returns(new DebugViewSnapshot(
+                "inst-a",
+                Array.Empty<AttributeValueChanged>(),
+                new[] { new AlarmStateChanged("inst-a", "A-alarm", AlarmState.Active, 500, T0) },
+                T0));
+
+        var services = new ServiceCollection();
+        services.AddSingleton(_instanceRepo);
+        services.AddSingleton(_siteRepo);
+        services.AddSingleton(_snapshotClient);
+        services.AddScoped<AlarmSummaryService>();
+        _provider = services.BuildServiceProvider();
+    }
+
+    private SharedAlarmSummaryService CreateSut(TimeSpan liveCacheTtl) =>
+        new(_provider.GetRequiredService<IServiceScopeFactory>(), _liveCache, liveCacheTtl, () => _now);
+
+    [Fact]
+    public async Task ConcurrentCircuits_ShareOneFanOut()
+    {
+        var sut = CreateSut(TimeSpan.FromSeconds(60));
+
+        var results = await Task.WhenAll(Enumerable.Range(0, 8).Select(_ => sut.GetSiteAlarmsAsync(SiteId)));
+
+        await _instanceRepo.Received(1).GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>());
+        Assert.All(results, r => Assert.Single(r.Alarms));
+    }
+
+    [Fact]
+    public async Task ColdLiveCache_RefreshesWithinThePageTick()
+    {
+        var sut = CreateSut(TimeSpan.FromSeconds(60));
+        _liveCache.Live = false;
+
+        await sut.GetSiteAlarmsAsync(SiteId);
+        // The page polls every 15s and, while the cache is cold, the poll is its
+        // authoritative rebuild — so the memo must have expired by then.
+        _now = T0.AddSeconds(15);
+        await sut.GetSiteAlarmsAsync(SiteId);
+
+        await _instanceRepo.Received(2).GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>());
+    }
+
+    [Fact]
+    public async Task LiveCacheServing_WidensTheWindowToTheReconcileInterval()
+    {
+        var sut = CreateSut(TimeSpan.FromSeconds(60));
+        _liveCache.Live = true;
+
+        await sut.GetSiteAlarmsAsync(SiteId);
+        _now = T0.AddSeconds(30);
+        await sut.GetSiteAlarmsAsync(SiteId);
+
+        // Live deltas own the rows; only the not-reporting list still comes from the
+        // fan-out, so a 30s-old answer is fine and costs no second fan-out.
+        await _instanceRepo.Received(1).GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>());
+
+        _now = T0.AddSeconds(61);
+        await sut.GetSiteAlarmsAsync(SiteId);
+        await _instanceRepo.Received(2).GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>());
+    }
+
+    [Fact]
+    public async Task DifferentSites_DoNotShareAMemoSlot()
+    {
+        const int otherSite = 8;
+        _siteRepo.GetSiteByIdAsync(otherSite, Arg.Any<CancellationToken>())
+            .Returns(new Site("Plant B", "plant-b") { Id = otherSite });
+        _instanceRepo.GetInstancesBySiteIdAsync(otherSite, Arg.Any<CancellationToken>())
+            .Returns(new List<Instance>());
+
+        var sut = CreateSut(TimeSpan.FromSeconds(60));
+
+        await sut.GetSiteAlarmsAsync(SiteId);
+        await sut.GetSiteAlarmsAsync(otherSite);
+
+        await _instanceRepo.Received(1).GetInstancesBySiteIdAsync(SiteId, Arg.Any<CancellationToken>());
+        await _instanceRepo.Received(1).GetInstancesBySiteIdAsync(otherSite, Arg.Any<CancellationToken>());
+    }
+
+    [Fact]
+    public void PureMethods_MatchTheDirectImplementation()
+    {
+        var sut = CreateSut(TimeSpan.FromSeconds(60));
+        var direct = new AlarmSummaryService(_instanceRepo, _siteRepo, _snapshotClient);
+        var alarms = new List<AlarmStateChanged>
+        {
+            new("inst-b", "B-alarm", AlarmState.Active, 900, T0),
+            new("inst-a", "A-alarm", AlarmState.Normal, 100, T0),
+        };
+
+        var shared = sut.BuildFromLiveAlarms(alarms);
+        var expected = direct.BuildFromLiveAlarms(alarms);
+
+        Assert.Equal(
+            expected.Alarms.Select(r => r.Alarm.AlarmName),
+            shared.Alarms.Select(r => r.Alarm.AlarmName));
+        var expectedRollup = direct.ComputeRollup(expected.Alarms);
+        var sharedRollup = sut.ComputeRollup(shared.Alarms);
+        Assert.Equal(expectedRollup.TotalActive, sharedRollup.TotalActive);
+        Assert.Equal(expectedRollup.WorstSeverity, sharedRollup.WorstSeverity);
+        Assert.Equal(expectedRollup.UnackedCount, sharedRollup.UnackedCount);
+        Assert.Equal(expectedRollup.CountsByKind, sharedRollup.CountsByKind);
+    }
+
+    public void Dispose() => _provider.Dispose();
+
+    /// <summary>Liveness-only stub — the façade consults nothing else on the live cache.</summary>
+    private sealed class FakeLiveCache : ISiteAlarmLiveCache
+    {
+        public bool Live { get; set; }
+
+        public IDisposable Subscribe(int siteId, Action onChanged) => new NoOp();
+
+        public IReadOnlyList<AlarmStateChanged> GetCurrentAlarms(int siteId) =>
+            Array.Empty<AlarmStateChanged>();
+
+        public bool IsLive(int siteId) => Live;
+
+        private sealed class NoOp : IDisposable
+        {
+            public void Dispose() { }
+        }
+    }
+}

tests/ZB.MOM.WW.ScadaBridge.CentralUI.Tests/Services/SingleFlightMemoTests.cs

@@ -0,0 +1,166 @@
+using ZB.MOM.WW.ScadaBridge.CentralUI.Services;
+
+namespace ZB.MOM.WW.ScadaBridge.CentralUI.Tests.Services;
+
+/// <summary>
+/// Unit tests for <see cref="SingleFlightMemo{T}"/> — the primitive behind the process-level
+/// KPI snapshot cache and the shared Alarm Summary fan-out (arch-review WP2.4). The two
+/// properties the whole optimisation rests on are tested here: <b>single-flight</b> (N
+/// concurrent circuits asking the same question produce ONE query) and <b>TTL</b> (a query
+/// answered inside the freshness window is served from memory, and one outside it is not).
+/// </summary>
+public class SingleFlightMemoTests
+{
+    private static readonly DateTimeOffset T0 = new(2026, 8, 14, 9, 0, 0, TimeSpan.Zero);
+
+    [Fact]
+    public async Task ConcurrentCallers_ShareOneFlight()
+    {
+        var released = new TaskCompletionSource<int>(TaskCreationOptions.RunContinuationsAsynchronously);
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(10), () => T0);
+
+        // Ten circuits ask while the first flight is still in progress.
+        var waiters = Enumerable.Range(0, 10)
+            .Select(_ => memo.GetAsync(() =>
+            {
+                Interlocked.Increment(ref invocations);
+                return released.Task;
+            }))
+            .ToArray();
+
+        released.SetResult(42);
+        var results = await Task.WhenAll(waiters);
+
+        Assert.Equal(1, invocations);
+        Assert.Equal(1, memo.FlightCount);
+        Assert.All(results, r => Assert.Equal(42, r));
+    }
+
+    [Fact]
+    public async Task WithinTtl_ServesTheMemoizedValue_WithoutRequerying()
+    {
+        var now = T0;
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(8), () => now);
+
+        var first = await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        now = T0.AddSeconds(7);
+        var second = await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        Assert.Equal(1, first);
+        Assert.Equal(1, second);
+        Assert.Equal(1, memo.FlightCount);
+    }
+
+    [Fact]
+    public async Task AfterTtl_RequeriesOnce()
+    {
+        var now = T0;
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(8), () => now);
+
+        await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        now = T0.AddSeconds(9);
+        var refreshed = await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        Assert.Equal(2, refreshed);
+        Assert.Equal(2, memo.FlightCount);
+    }
+
+    [Fact]
+    public async Task ForceRefresh_BypassesAFreshValue()
+    {
+        var now = T0;
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(60), () => now);
+
+        await memo.GetAsync(() => Task.FromResult(++invocations));
+        var forced = await memo.GetAsync(() => Task.FromResult(++invocations), forceRefresh: true);
+
+        Assert.Equal(2, forced);
+        Assert.Equal(2, memo.FlightCount);
+    }
+
+    [Fact]
+    public async Task ForceRefresh_StillJoinsAnInFlightRound()
+    {
+        var released = new TaskCompletionSource<int>(TaskCreationOptions.RunContinuationsAsynchronously);
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(60), () => T0);
+
+        var first = memo.GetAsync(() =>
+        {
+            Interlocked.Increment(ref invocations);
+            return released.Task;
+        });
+        // A burst of operator "Refresh" clicks must not multiply the query.
+        var forcedA = memo.GetAsync(() => Task.FromResult(-1), forceRefresh: true);
+        var forcedB = memo.GetAsync(() => Task.FromResult(-1), forceRefresh: true);
+
+        released.SetResult(7);
+        Assert.Equal(7, await first);
+        Assert.Equal(7, await forcedA);
+        Assert.Equal(7, await forcedB);
+        Assert.Equal(1, invocations);
+    }
+
+    [Fact]
+    public async Task Failure_IsNotMemoized_AndTheNextCallerRetries()
+    {
+        var now = T0;
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(60), () => now);
+
+        await Assert.ThrowsAsync<InvalidOperationException>(() =>
+            memo.GetAsync(() =>
+            {
+                invocations++;
+                return Task.FromException<int>(new InvalidOperationException("KPI query failed"));
+            }));
+
+        // Same instant, well inside the TTL — a cached fault would starve the tiles for a
+        // whole window, so the next caller must produce a new flight.
+        var recovered = await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        Assert.Equal(2, recovered);
+        Assert.Equal(2, memo.FlightCount);
+    }
+
+    [Fact]
+    public async Task TtlOverride_AppliesToThatFlight()
+    {
+        var now = T0;
+        var invocations = 0;
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(5), () => now);
+
+        // Produced under a 60s override — still fresh 30s later despite the 5s default.
+        await memo.GetAsync(() => Task.FromResult(++invocations), ttlOverride: TimeSpan.FromSeconds(60));
+
+        now = T0.AddSeconds(30);
+        var second = await memo.GetAsync(() => Task.FromResult(++invocations));
+
+        Assert.Equal(1, second);
+        Assert.Equal(1, memo.FlightCount);
+    }
+
+    [Fact]
+    public async Task CallerCancellation_DoesNotCancelTheSharedFlight()
+    {
+        var released = new TaskCompletionSource<int>(TaskCreationOptions.RunContinuationsAsynchronously);
+        var memo = new SingleFlightMemo<int>(TimeSpan.FromSeconds(60), () => T0);
+
+        using var giveUp = new CancellationTokenSource();
+        var impatient = memo.GetAsync(() => released.Task, cancellationToken: giveUp.Token);
+        var patient = memo.GetAsync(() => Task.FromResult(-1));
+
+        giveUp.Cancel();
+        await Assert.ThrowsAnyAsync<OperationCanceledException>(() => impatient);
+
+        // The abandoned wait must not have taken the round down with it.
+        released.SetResult(99);
+        Assert.Equal(99, await patient);
+    }
+}