Merge branch 'feature/audit-log-m6-reconciliation-purge': Audit Log #23 M6 Reconciliation + Purge + Partition Maintenance + Health Metrics

M6 ships the self-healing + lifecycle-maintenance layer:
- PullAuditEvents RPC + site-side handler (sitestream.proto extended;
  ISiteAuditQueue.ReadPendingSinceAsync + MarkReconciledAsync).
- SiteAuditReconciliationActor central singleton: per-site 5-min cursor,
  pulls via mockable IPullAuditEventsClient seam (real gRPC client deferred
  to a follow-up), ingests via existing AuditLogIngestActor path.
- AuditLogPurgeActor + repository fix: SwitchOutPartitionAsync replaced
  with DROP INDEX → SWITCH PARTITION → DROP staging → CREATE INDEX dance.
  M1 NotSupportedException stub retired.
- AuditLogPartitionMaintenanceService IHostedService: monthly SPLIT
  RANGE roll-forward; explicit ALTER PARTITION SCHEME NEXT USED before
  each SPLIT (critical fix — ALL TO PRIMARY auto-populates NEXT USED
  only on the first SPLIT).
- Health metrics: SiteAuditBacklog (count + age + bytes) per site;
  SiteAuditTelemetryStalledTracker subscribes to EventStream;
  CentralAuditWriteFailures counter + IAuditCentralHealthSnapshot
  aggregator; central-side AuditRedactionFailure routed to the snapshot.

Site-side AuditRedactionFailure counter (M5) stays on the site bridge;
central uses the new CentralAuditRedactionFailureCounter.

Integration tests: outage+reconciliation (200 events buffered, drained
on recovery, no duplicates), partition-switch purge (drop-and-rebuild
verified), partition maintenance roll-forward (idempotent across two
ticks).

Shipped: 16 commits, ~75 net new tests. Full solution 24/24 test
projects green. Pre-existing M5-era Host.Tests CTS-disposal flake
incidentally fixed by the Bundle E lock-guarded _trackedDisposables
enumeration.

Latent EventStream-after-await bug in SiteAuditReconciliationActor
fixed in Bundle D (D0). Production wiring of real gRPC pull client
deferred to a follow-up bundle — actor + abstractions are testable
via the mockable seam. infra/* untouched on any branch commit.

docs/plans/2026-05-20-auditlog-m6-reconciliation-purge.md

@@ -0,0 +1,19 @@
+# Audit Log #23 — M6 Reconciliation + Purge + Partition Maintenance + Health Metrics
+
+> **For Claude:** subagent-driven-development with bundled cadence.
+
+**Goal:** Self-healing telemetry (5-min reconciliation pull), monthly partition rollover, daily partition-switch purge with drop-and-rebuild around UX_AuditLog_EventId, all five health metrics live (SiteAuditBacklog, SiteAuditWriteFailures, SiteAuditTelemetryStalled, CentralAuditWriteFailures, AuditRedactionFailure).
+
+**M5 realities baked in:** AuditRedactionFailure counter is site-only — M6-T9 surfaces it centrally. SwitchOutPartitionAsync ships as NotSupportedException stub from M1; M6-T4 replaces it with the drop-DROP-INDEX → SWITCH PARTITION → DROP staging → CREATE UNIQUE NONCLUSTERED INDEX dance. Partition function pre-seeded Jan 2026 – Dec 2027; M6-T5 SPLITs new boundaries forward.
+
+**Bundles:**
+- Bundle A — Proto + site handler (T1, T2)
+- Bundle B — Reconciliation actor (T3)
+- Bundle C — Purge actor + drop-and-rebuild repository fix (T4)
+- Bundle D — Partition maintenance hosted service (T5)
+- Bundle E — Health metrics (T6, T7, T8, T9)
+- Bundle F — Integration tests (T10, T11, T12)
+
+Final cross-bundle review + merge.
+
+**Note**: M2 noted NoOpSiteStreamAuditClient stays in production until "M6 wires the real client". M6-T1+T2 add the PULL RPC; the actual production PUSH client (real implementation of ISiteStreamAuditClient.IngestAuditEventsAsync + IngestCachedTelemetryAsync) is the bigger lift. M6 will add the real client IF feasible within scope OR defer to a follow-up. Decision: try in Bundle A (alongside the proto extension); if scope blows up, the NoOp stays.

src/ScadaLink.AuditLog/Central/AuditCentralHealthSnapshot.cs

@@ -0,0 +1,80 @@
+using System.Collections.Concurrent;
+using ScadaLink.AuditLog.Payload;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T8, T9) — central singleton implementation of
+/// <see cref="IAuditCentralHealthSnapshot"/>. Owns thread-safe
+/// <see cref="System.Threading.Interlocked"/> counters for
+/// <c>CentralAuditWriteFailures</c> + <c>AuditRedactionFailure</c> and a
+/// per-site latched stalled-state map fed by the
+/// <see cref="SiteAuditTelemetryStalledTracker"/>. Also implements the
+/// writer surfaces (<see cref="ICentralAuditWriteFailureCounter"/> +
+/// <see cref="IAuditRedactionFailureCounter"/>) so a single concrete object
+/// is the source of truth — DI binds those two interfaces to this same
+/// singleton instance on the central composition root.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Why one type for read + write.</b> The writer interfaces are tiny
+/// (<c>Increment()</c>) and the read surface needs visibility of those
+/// counters anyway — having a single class own both means the
+/// <c>Interlocked</c> field IS the snapshot value, no extra plumbing needed.
+/// Mirrors the
+/// <see cref="ScadaLink.HealthMonitoring.SiteHealthCollector"/> pattern where
+/// the collector both receives and exposes the metric.
+/// </para>
+/// <para>
+/// <b>Stalled-state plumbing.</b> The per-site stalled latch lives directly
+/// on this snapshot. <see cref="SiteAuditTelemetryStalledTracker"/> is the
+/// EventStream subscriber that pushes
+/// <see cref="SiteAuditTelemetryStalledChanged"/> publications in via
+/// <see cref="ApplyStalled"/>. Keeping the dictionary on this type (rather
+/// than reading the tracker on every access) lets the snapshot be constructed
+/// without an <see cref="Akka.Actor.ActorSystem"/> dependency — the tracker
+/// is wired up later from the Akka bootstrap, once the system is built.
+/// </para>
+/// </remarks>
+public sealed class AuditCentralHealthSnapshot
+    : IAuditCentralHealthSnapshot,
+      ICentralAuditWriteFailureCounter,
+      IAuditRedactionFailureCounter
+{
+    private int _centralAuditWriteFailures;
+    private int _auditRedactionFailure;
+    private readonly ConcurrentDictionary<string, bool> _stalled = new();
+
+    /// <inheritdoc/>
+    public int CentralAuditWriteFailures =>
+        Interlocked.CompareExchange(ref _centralAuditWriteFailures, 0, 0);
+
+    /// <inheritdoc/>
+    public int AuditRedactionFailure =>
+        Interlocked.CompareExchange(ref _auditRedactionFailure, 0, 0);
+
+    /// <inheritdoc/>
+    public IReadOnlyDictionary<string, bool> SiteAuditTelemetryStalled =>
+        new Dictionary<string, bool>(_stalled);
+
+    /// <summary>
+    /// Apply a <see cref="SiteAuditTelemetryStalledChanged"/> publication
+    /// observed by <see cref="SiteAuditTelemetryStalledTracker"/>. Public
+    /// so the tracker (which lives in the same assembly but is constructed
+    /// later from the Akka host) can push without a friend reference;
+    /// readers should call <see cref="SiteAuditTelemetryStalled"/>.
+    /// </summary>
+    public void ApplyStalled(SiteAuditTelemetryStalledChanged evt)
+    {
+        if (evt is null) return;
+        _stalled[evt.SiteId] = evt.Stalled;
+    }
+
+    /// <inheritdoc/>
+    void ICentralAuditWriteFailureCounter.Increment() =>
+        Interlocked.Increment(ref _centralAuditWriteFailures);
+
+    /// <inheritdoc/>
+    void IAuditRedactionFailureCounter.Increment() =>
+        Interlocked.Increment(ref _auditRedactionFailure);
+}

src/ScadaLink.AuditLog/Central/AuditLogIngestActor.cs

@@ -124,6 +124,7 @@ public class AuditLogIngestActor : ReceiveActor
         IServiceScope? scope = null;
         IAuditLogRepository repository;
         IAuditPayloadFilter? filter = null;
+        ICentralAuditWriteFailureCounter? failureCounter = null;
         if (_injectedRepository is not null)
         {
             repository = _injectedRepository;
@@ -133,6 +134,10 @@ public class AuditLogIngestActor : ReceiveActor
             scope = _serviceProvider!.CreateScope();
             repository = scope.ServiceProvider.GetRequiredService<IAuditLogRepository>();
             filter = scope.ServiceProvider.GetService<IAuditPayloadFilter>();
+            // M6 Bundle E (T8): central health counter is best-effort —
+            // unregistered (test composition roots) means the per-row catch
+            // simply logs without surfacing on the health dashboard.
+            failureCounter = scope.ServiceProvider.GetService<ICentralAuditWriteFailureCounter>();
         }
 
         try
@@ -157,6 +162,10 @@ public class AuditLogIngestActor : ReceiveActor
                 {
                     // Per-row catch — one bad row never sinks the whole batch.
                     // The row stays Pending at the site; the next drain retries.
+                    // M6 Bundle E (T8): bump the central health counter so a
+                    // sustained insert-throw failure surfaces on the dashboard.
+                    try { failureCounter?.Increment(); }
+                    catch { /* counter must never throw — defence in depth */ }
                     _logger.LogError(ex,
                         "Failed to persist audit event {EventId} during batch ingest; row will be retried by the site.",
                         evt.EventId);
@@ -204,6 +213,10 @@ public class AuditLogIngestActor : ReceiveActor
             // never throw, so we can apply it inside the per-entry try
             // without risking an unbounded blast radius.
             var filter = scope.ServiceProvider.GetService<IAuditPayloadFilter>();
+            // M6 Bundle E (T8): same best-effort central health counter as
+            // the OnIngestAsync path — null on test composition roots that
+            // skip the registration.
+            var failureCounter = scope.ServiceProvider.GetService<ICentralAuditWriteFailureCounter>();
 
             foreach (var entry in cmd.Entries)
             {
@@ -240,6 +253,10 @@ public class AuditLogIngestActor : ReceiveActor
                     // EventId is NOT added to `accepted` so the site keeps its
                     // row Pending and retries on the next drain. Other entries
                     // in the batch continue with their own transactions.
+                    // M6 Bundle E (T8): bump the central health counter so a
+                    // sustained dual-write failure surfaces on the dashboard.
+                    try { failureCounter?.Increment(); }
+                    catch { /* counter must never throw — defence in depth */ }
                     _logger.LogError(
                         ex,
                         "Combined telemetry dual-write failed for AuditEvent {EventId} / TrackedOperationId {TrackedOpId}; rolled back.",

src/ScadaLink.AuditLog/Central/AuditLogPartitionMaintenanceOptions.cs

@@ -0,0 +1,37 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Tuning knobs for the central
+/// <see cref="AuditLogPartitionMaintenanceService"/> hosted service (M6-T5).
+/// Defaults: once every 24 hours, keep at least one future monthly
+/// boundary ahead of <see cref="DateTime.UtcNow"/>.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The hosted service drives a daily roll-forward of
+/// <c>pf_AuditLog_Month</c>: each tick reads the current max boundary and
+/// SPLITs new monthly boundaries until at least
+/// <see cref="LookaheadMonths"/> future months are covered. The 1-month
+/// default is intentionally conservative — anything less risks an
+/// end-of-month race where inserts land in the unbounded tail partition;
+/// anything more wastes nothing but represents premature commitment.
+/// </para>
+/// <para>
+/// The 24-hour cadence is the cheapest interval that still guarantees
+/// at-most-one missed boundary in steady state (even a hard failover the
+/// hosted service can recover on its very next tick). Lowering this below
+/// an hour would generate more metadata churn than it saves.
+/// </para>
+/// </remarks>
+public sealed class AuditLogPartitionMaintenanceOptions
+{
+    /// <summary>Period of the maintenance tick in seconds (default 86 400 = 24 h).</summary>
+    public int IntervalSeconds { get; set; } = 86_400;
+
+    /// <summary>
+    /// Minimum number of future months that <c>pf_AuditLog_Month</c> must
+    /// cover after each tick. Default 1 — i.e. as of mid-May the partition
+    /// for the next full month (June) must already be present.
+    /// </summary>
+    public int LookaheadMonths { get; set; } = 1;
+}

src/ScadaLink.AuditLog/Central/AuditLogPartitionMaintenanceService.cs

@@ -0,0 +1,145 @@
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Hosting;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Options;
+using ScadaLink.Commons.Interfaces;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Central <see cref="IHostedService"/> (M6-T5, Bundle D) that rolls
+/// <c>pf_AuditLog_Month</c> forward once a day. Each tick opens a fresh DI
+/// scope, resolves <see cref="IPartitionMaintenance"/>, and calls
+/// <see cref="IPartitionMaintenance.EnsureLookaheadAsync"/> to SPLIT any
+/// missing future boundaries — the partition function must always cover at
+/// least <see cref="AuditLogPartitionMaintenanceOptions.LookaheadMonths"/>
+/// future months, otherwise inserts past the highest boundary accumulate in
+/// a single unbounded tail partition that <c>SwitchOutPartitionAsync</c>
+/// cannot purge cleanly.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Why a hosted service, not an actor.</b> Bundle C's
+/// <see cref="AuditLogPurgeActor"/> sits inside the central singleton
+/// because it needs supervised lifecycle alongside the rest of the
+/// reconciliation / ingest pipeline. Roll-forward is genuinely a once-a-day
+/// chore with no cross-actor coordination, so we use the much simpler
+/// hosted-service pattern: <c>Task.Run</c> on start, <c>Task.Delay</c>
+/// between ticks, cancellation on stop. Reusing
+/// <see cref="IPartitionMaintenance"/> from the central node-only DI graph
+/// keeps the contract testable without any actor framework involvement.
+/// </para>
+/// <para>
+/// <b>Failure containment.</b> The tick body wraps the maintenance call in
+/// a try/catch so a transient SQL Server error never tears down the hosted
+/// service — the next tick simply retries. The exception is logged with
+/// the original stack trace at <c>Error</c> level; ops surfaces (M6 Bundle
+/// E's central health collector) can subscribe to the logger to alert on
+/// repeated failures.
+/// </para>
+/// <para>
+/// <b>Startup ordering.</b> A first tick fires immediately at
+/// <see cref="StartAsync"/> so a fresh deployment doesn't need to wait
+/// <see cref="AuditLogPartitionMaintenanceOptions.IntervalSeconds"/> for
+/// the partition function to come up to spec. This is also what the brief
+/// asks for ("Run once on startup").
+/// </para>
+/// <para>
+/// <b>DI scope per tick.</b> <see cref="IPartitionMaintenance"/> is scoped
+/// (alongside the rest of the EF repositories) because the implementation
+/// reuses the per-scope <c>ScadaLinkDbContext</c>. A hosted service is a
+/// singleton, so it must open and dispose a scope around each tick — the
+/// same pattern <see cref="AuditLogPurgeActor"/> uses.
+/// </para>
+/// </remarks>
+public sealed class AuditLogPartitionMaintenanceService : IHostedService, IDisposable
+{
+    private readonly IServiceScopeFactory _scopeFactory;
+    private readonly IOptions<AuditLogPartitionMaintenanceOptions> _options;
+    private readonly ILogger<AuditLogPartitionMaintenanceService> _logger;
+    private CancellationTokenSource? _cts;
+    private Task? _loop;
+
+    public AuditLogPartitionMaintenanceService(
+        IServiceScopeFactory scopeFactory,
+        IOptions<AuditLogPartitionMaintenanceOptions> options,
+        ILogger<AuditLogPartitionMaintenanceService> logger)
+    {
+        _scopeFactory = scopeFactory ?? throw new ArgumentNullException(nameof(scopeFactory));
+        _options = options ?? throw new ArgumentNullException(nameof(options));
+        _logger = logger ?? throw new ArgumentNullException(nameof(logger));
+    }
+
+    /// <inheritdoc />
+    public Task StartAsync(CancellationToken ct)
+    {
+        // Linked CTS lets StopAsync's cancellation AND the host's shutdown
+        // token both terminate the loop; either side firing aborts the
+        // pending Task.Delay.
+        _cts = CancellationTokenSource.CreateLinkedTokenSource(ct);
+        _loop = Task.Run(() => RunLoopAsync(_cts.Token));
+        return Task.CompletedTask;
+    }
+
+    private async Task RunLoopAsync(CancellationToken ct)
+    {
+        // Run once on startup so a fresh deployment isn't gated on the
+        // IntervalSeconds initial wait — the brief calls this out explicitly.
+        await SafeMaintainAsync(ct).ConfigureAwait(false);
+
+        while (!ct.IsCancellationRequested)
+        {
+            try
+            {
+                await Task.Delay(TimeSpan.FromSeconds(_options.Value.IntervalSeconds), ct)
+                    .ConfigureAwait(false);
+            }
+            catch (OperationCanceledException)
+            {
+                break;
+            }
+
+            await SafeMaintainAsync(ct).ConfigureAwait(false);
+        }
+    }
+
+    private async Task SafeMaintainAsync(CancellationToken ct)
+    {
+        try
+        {
+            await using var scope = _scopeFactory.CreateAsyncScope();
+            var maintenance = scope.ServiceProvider.GetRequiredService<IPartitionMaintenance>();
+            var added = await maintenance
+                .EnsureLookaheadAsync(_options.Value.LookaheadMonths, ct)
+                .ConfigureAwait(false);
+            if (added.Count > 0)
+            {
+                _logger.LogInformation(
+                    "AuditLogPartitionMaintenance added {Count} boundaries: {Boundaries}",
+                    added.Count,
+                    string.Join(", ", added.Select(b => b.ToString("yyyy-MM-dd"))));
+            }
+        }
+        catch (Exception ex)
+        {
+            // Catch-all is deliberate: the hosted service must survive every
+            // class of tick failure (transient SQL, DI resolution, etc.) so
+            // the next tick gets a chance. The brief's contract is
+            // "exception logged, not propagated".
+            _logger.LogError(ex, "AuditLogPartitionMaintenance tick failed");
+        }
+    }
+
+    /// <inheritdoc />
+    public Task StopAsync(CancellationToken ct)
+    {
+        _cts?.Cancel();
+        return _loop ?? Task.CompletedTask;
+    }
+
+    /// <inheritdoc />
+    public void Dispose()
+    {
+        _cts?.Dispose();
+    }
+}

src/ScadaLink.AuditLog/Central/AuditLogPurgeActor.cs

@@ -0,0 +1,214 @@
+using System.Diagnostics;
+using Akka.Actor;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Configuration;
+using ScadaLink.Commons.Interfaces.Repositories;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Central singleton (M6 Bundle C) that drives the daily AuditLog partition
+/// purge. On a configurable timer (default 24 hours) the actor:
+/// <list type="number">
+/// <item>Queries <see cref="IAuditLogRepository.GetPartitionBoundariesOlderThanAsync"/>
+///       for monthly boundaries whose latest <c>OccurredAtUtc</c> is older
+///       than <c>DateTime.UtcNow - RetentionDays</c>.</item>
+/// <item>For each eligible boundary, calls
+///       <see cref="IAuditLogRepository.SwitchOutPartitionAsync"/> which runs
+///       the drop-and-rebuild dance around <c>UX_AuditLog_EventId</c>.</item>
+/// <item>Publishes <see cref="AuditLogPurgedEvent"/> on the actor-system
+///       EventStream so the Bundle E central health collector + ops surfaces
+///       can subscribe without coupling to this actor.</item>
+/// </list>
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Daily cadence.</b> Partition switch is metadata-only but the
+/// drop-and-rebuild dance briefly removes <c>UX_AuditLog_EventId</c>; running
+/// more often than necessary trades unique-index rebuild outages for
+/// negligible freshness wins. The default 24-hour interval matches
+/// alog.md §10's retention policy.
+/// </para>
+/// <para>
+/// <b>Continue-on-error.</b> A single boundary that throws (transient SQL
+/// failure, contention with backup, missing object) must NOT prevent the
+/// other eligible boundaries from being purged on the same tick. Per-boundary
+/// work runs inside its own try/catch; the actor's
+/// <see cref="SupervisorStrategy"/> uses Resume so any leaked exception keeps
+/// the singleton alive for the next tick.
+/// </para>
+/// <para>
+/// <b>DI scopes.</b> <see cref="IAuditLogRepository"/> is a scoped EF Core
+/// service registered by <c>AddConfigurationDatabase</c>. The singleton
+/// opens one DI scope per tick and reuses the same repository across every
+/// boundary in that tick — mirrors the
+/// <see cref="SiteAuditReconciliationActor"/> pattern.
+/// </para>
+/// <para>
+/// <b>EventStream.</b> Publishing <see cref="AuditLogPurgedEvent"/> through
+/// the EventStream rather than direct messaging avoids coupling this actor
+/// to its consumers; M6 Bundle E will subscribe a central health-counter
+/// bridge that surfaces purge progress on the central health report.
+/// </para>
+/// </remarks>
+public class AuditLogPurgeActor : ReceiveActor
+{
+    private readonly IServiceProvider _services;
+    private readonly AuditLogPurgeOptions _purgeOptions;
+    private readonly AuditLogOptions _auditOptions;
+    private readonly ILogger<AuditLogPurgeActor> _logger;
+    private ICancelable? _timer;
+
+    public AuditLogPurgeActor(
+        IServiceProvider services,
+        IOptions<AuditLogPurgeOptions> purgeOptions,
+        IOptions<AuditLogOptions> auditOptions,
+        ILogger<AuditLogPurgeActor> logger)
+    {
+        ArgumentNullException.ThrowIfNull(services);
+        ArgumentNullException.ThrowIfNull(purgeOptions);
+        ArgumentNullException.ThrowIfNull(auditOptions);
+        ArgumentNullException.ThrowIfNull(logger);
+
+        _services = services;
+        _purgeOptions = purgeOptions.Value;
+        _auditOptions = auditOptions.Value;
+        _logger = logger;
+
+        ReceiveAsync<PurgeTick>(_ => OnTickAsync());
+    }
+
+    protected override void PreStart()
+    {
+        base.PreStart();
+        var interval = _purgeOptions.Interval;
+        _timer = Context.System.Scheduler.ScheduleTellRepeatedlyCancelable(
+            initialDelay: interval,
+            interval: interval,
+            receiver: Self,
+            message: PurgeTick.Instance,
+            sender: Self);
+    }
+
+    protected override void PostStop()
+    {
+        _timer?.Cancel();
+        base.PostStop();
+    }
+
+    /// <summary>
+    /// Resume keeps the singleton alive across any leaked exception. Restart
+    /// would re-run PreStart and reschedule the timer (harmless but wasteful);
+    /// Stop is wrong because the singleton must keep ticking until shutdown.
+    /// </summary>
+    protected override SupervisorStrategy SupervisorStrategy()
+    {
+        return new OneForOneStrategy(
+            maxNrOfRetries: 0,
+            withinTimeRange: TimeSpan.Zero,
+            decider: Akka.Actor.SupervisorStrategy.DefaultDecider);
+    }
+
+    private async Task OnTickAsync()
+    {
+        // Capture EventStream BEFORE the first await. Accessing Context (and
+        // therefore Context.System) after an await is unsafe because Akka's
+        // ActorBase.Context throws "no active ActorContext" once the
+        // continuation runs on a thread that isn't currently dispatching this
+        // actor — mirrors the same Sender-capture pattern in
+        // AuditLogIngestActor.OnIngestAsync.
+        var eventStream = Context.System.EventStream;
+
+        // Compute the retention threshold from AuditLogOptions.RetentionDays
+        // each tick — the options class supports hot reload via
+        // IOptionsMonitor for the redaction policy and similar settings; we
+        // read the snapshot per-tick so an operator who lowers RetentionDays
+        // sees the change applied on the next purge without an actor
+        // restart.
+        var threshold = DateTime.UtcNow - TimeSpan.FromDays(_auditOptions.RetentionDays);
+
+        IServiceScope? scope = null;
+        IAuditLogRepository repository;
+        try
+        {
+            scope = _services.CreateScope();
+            repository = scope.ServiceProvider.GetRequiredService<IAuditLogRepository>();
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex, "Failed to resolve IAuditLogRepository for AuditLog purge tick.");
+            scope?.Dispose();
+            return;
+        }
+
+        try
+        {
+            IReadOnlyList<DateTime> boundaries;
+            try
+            {
+                boundaries = await repository
+                    .GetPartitionBoundariesOlderThanAsync(threshold)
+                    .ConfigureAwait(false);
+            }
+            catch (Exception ex)
+            {
+                _logger.LogError(
+                    ex,
+                    "Failed to enumerate eligible AuditLog partition boundaries (threshold {ThresholdUtc:o}); skipping purge tick.",
+                    threshold);
+                return;
+            }
+
+            if (boundaries.Count == 0)
+            {
+                return;
+            }
+
+            foreach (var boundary in boundaries)
+            {
+                // Per-boundary try/catch: one bad partition (transient SQL
+                // failure, missing object, contention with backup) does NOT
+                // abandon the rest of the tick.
+                var sw = Stopwatch.StartNew();
+                try
+                {
+                    var rowsDeleted = await repository
+                        .SwitchOutPartitionAsync(boundary)
+                        .ConfigureAwait(false);
+                    sw.Stop();
+
+                    eventStream.Publish(
+                        new AuditLogPurgedEvent(boundary, rowsDeleted, sw.ElapsedMilliseconds));
+
+                    _logger.LogInformation(
+                        "Purged AuditLog partition {MonthBoundary:yyyy-MM-dd}; {RowsDeleted} rows in {DurationMs} ms.",
+                        boundary,
+                        rowsDeleted,
+                        sw.ElapsedMilliseconds);
+                }
+                catch (Exception ex)
+                {
+                    sw.Stop();
+                    _logger.LogError(
+                        ex,
+                        "Failed to purge AuditLog partition {MonthBoundary:yyyy-MM-dd}; other partitions continue. Elapsed {DurationMs} ms.",
+                        boundary,
+                        sw.ElapsedMilliseconds);
+                }
+            }
+        }
+        finally
+        {
+            scope.Dispose();
+        }
+    }
+
+    /// <summary>Self-tick triggering a purge pass across all eligible partitions.</summary>
+    internal sealed class PurgeTick
+    {
+        public static readonly PurgeTick Instance = new();
+        private PurgeTick() { }
+    }
+}

src/ScadaLink.AuditLog/Central/AuditLogPurgeOptions.cs

@@ -0,0 +1,43 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Tuning knobs for the central <see cref="AuditLogPurgeActor"/> singleton.
+/// Default cadence is 24 hours per the M6 plan; the retention window itself
+/// is sourced from <see cref="ScadaLink.AuditLog.Configuration.AuditLogOptions.RetentionDays"/>
+/// (default 365) so operators tune retention from a single section.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The purge actor is a daily-cadence singleton, not a hot-loop, because
+/// partition-switch I/O is metadata-only but the drop-and-rebuild dance
+/// briefly removes the <c>UX_AuditLog_EventId</c> unique index — running
+/// more often than necessary trades index-rebuild outages for marginal
+/// freshness gains. Lower this only when an operator can prove they need
+/// sub-daily purge granularity.
+/// </para>
+/// <para>
+/// <see cref="IntervalOverride"/> exists for tests to drop the cadence to
+/// milliseconds without polluting the production config surface; production
+/// binds <see cref="IntervalHours"/> only.
+/// </para>
+/// </remarks>
+public sealed class AuditLogPurgeOptions
+{
+    /// <summary>Period of the purge tick in hours (default 24).</summary>
+    public int IntervalHours { get; set; } = 24;
+
+    /// <summary>
+    /// Test-only override for finer control over the tick cadence than
+    /// whole-hour resolution allows. When non-null, takes precedence over
+    /// <see cref="IntervalHours"/>. Not bound from config — production
+    /// config exposes <see cref="IntervalHours"/> only.
+    /// </summary>
+    public TimeSpan? IntervalOverride { get; set; }
+
+    /// <summary>
+    /// Resolves the effective tick interval, honouring the test override
+    /// when set. Falls back to <see cref="IntervalHours"/>.
+    /// </summary>
+    public TimeSpan Interval =>
+        IntervalOverride ?? TimeSpan.FromHours(IntervalHours);
+}

src/ScadaLink.AuditLog/Central/AuditLogPurgedEvent.cs

@@ -0,0 +1,29 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Published on the actor-system EventStream by <see cref="AuditLogPurgeActor"/>
+/// after each successful partition switch-out. Downstream consumers (Bundle E
+/// central health collector, ops dashboards, audit trails) subscribe so a
+/// purge action is observable without the actor needing to know about any
+/// specific subscriber.
+/// </summary>
+/// <param name="MonthBoundary">
+/// The pf_AuditLog_Month lower-bound boundary that was switched out — i.e.
+/// the first instant of the purged month in UTC.
+/// </param>
+/// <param name="RowsDeleted">
+/// Approximate row count purged from the partition, sampled BEFORE the
+/// switch. Exact accounting would require a post-switch scan of the staging
+/// table, which the dance drops immediately, so this is the closest
+/// observable proxy. Zero is a valid value when the actor's enumerator
+/// included a partition the operator subsequently emptied by hand.
+/// </param>
+/// <param name="DurationMs">
+/// Wall-clock time spent inside <c>SwitchOutPartitionAsync</c> for this
+/// boundary, in milliseconds. Useful for spotting the rare slow purge
+/// without spinning up dedicated telemetry.
+/// </param>
+public sealed record AuditLogPurgedEvent(
+    DateTime MonthBoundary,
+    long RowsDeleted,
+    long DurationMs);

src/ScadaLink.AuditLog/Central/CentralAuditRedactionFailureCounter.cs

@@ -0,0 +1,57 @@
+using ScadaLink.AuditLog.Payload;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T9) — bridges
+/// <see cref="IAuditRedactionFailureCounter"/> (incremented by
+/// <see cref="DefaultAuditPayloadFilter"/> every time a header / body / SQL
+/// parameter redactor stage throws and the filter has to over-redact the
+/// offending field) into <see cref="AuditCentralHealthSnapshot"/> so the
+/// failure surfaces on the central health surface as
+/// <c>AuditCentralHealthSnapshot.AuditRedactionFailure</c>.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Site vs central.</b> M5 Bundle C wired the SITE-side bridge
+/// (<see cref="ScadaLink.AuditLog.Site.HealthMetricsAuditRedactionFailureCounter"/>),
+/// which routes increments into the site health report payload's
+/// <c>AuditRedactionFailure</c> field. That handles redactor failures on the
+/// site SQLite hot-path (FallbackAuditWriter). M6 Bundle E (T9) adds the
+/// MIRROR bridge here so the same payload filter — when it runs on the
+/// central <see cref="CentralAuditWriter"/> /
+/// <see cref="AuditLogIngestActor"/> paths — surfaces its failures on the
+/// central dashboard rather than disappearing into a NoOp.
+/// </para>
+/// <para>
+/// <b>Registration shape.</b> Site composition roots call
+/// <see cref="ServiceCollectionExtensions.AddAuditLogHealthMetricsBridge"/>,
+/// which overrides the binding with the site bridge. Central composition
+/// roots call <see cref="ServiceCollectionExtensions.AddAuditLogCentralMaintenance"/>,
+/// which overrides with this central bridge. A node never wears both hats —
+/// site and central are distinct host roles — so the two bridges never
+/// fight over the same binding at runtime.
+/// </para>
+/// <para>
+/// <b>Why not a thin wrapper around the snapshot directly?</b> The snapshot
+/// itself <i>could</i> be the bound implementation (it already implements
+/// <see cref="IAuditRedactionFailureCounter"/>), but a dedicated class makes
+/// the central-vs-site asymmetry explicit at the DI boundary — readers of
+/// <see cref="ServiceCollectionExtensions.AddAuditLogCentralMaintenance"/>
+/// see "site → site bridge, central → central bridge", matching the
+/// <see cref="ScadaLink.AuditLog.Site.HealthMetricsAuditRedactionFailureCounter"/>
+/// shape one-for-one.
+/// </para>
+/// </remarks>
+public sealed class CentralAuditRedactionFailureCounter : IAuditRedactionFailureCounter
+{
+    private readonly AuditCentralHealthSnapshot _snapshot;
+
+    public CentralAuditRedactionFailureCounter(AuditCentralHealthSnapshot snapshot)
+    {
+        _snapshot = snapshot ?? throw new ArgumentNullException(nameof(snapshot));
+    }
+
+    /// <inheritdoc/>
+    public void Increment() => ((IAuditRedactionFailureCounter)_snapshot).Increment();
+}

src/ScadaLink.AuditLog/Central/CentralAuditWriter.cs

@@ -42,6 +42,7 @@ public sealed class CentralAuditWriter : ICentralAuditWriter
     private readonly IServiceProvider _services;
     private readonly ILogger<CentralAuditWriter> _logger;
     private readonly IAuditPayloadFilter? _filter;
+    private readonly ICentralAuditWriteFailureCounter _failureCounter;
 
     /// <summary>
     /// Bundle C (M5-T6) — the central direct-write path used by the
@@ -50,15 +51,23 @@ public sealed class CentralAuditWriter : ICentralAuditWriter
     /// optional so the M4 test composition roots that don't pass one keep
     /// working (they only ever write small payloads); production DI registers
     /// the real filter via <see cref="ServiceCollectionExtensions.AddAuditLog"/>.
+    /// M6 Bundle E (T8) — adds the optional
+    /// <see cref="ICentralAuditWriteFailureCounter"/> so a swallowed repository
+    /// throw bumps the central health surface's
+    /// <c>CentralAuditWriteFailures</c> counter. Defaults to a NoOp so test
+    /// composition roots that don't wire the counter keep their current
+    /// behaviour.
     /// </summary>
     public CentralAuditWriter(
         IServiceProvider services,
         ILogger<CentralAuditWriter> logger,
-        IAuditPayloadFilter? filter = null)
+        IAuditPayloadFilter? filter = null,
+        ICentralAuditWriteFailureCounter? failureCounter = null)
     {
         _services = services ?? throw new ArgumentNullException(nameof(services));
         _logger = logger ?? throw new ArgumentNullException(nameof(logger));
         _filter = filter;
+        _failureCounter = failureCounter ?? new NoOpCentralAuditWriteFailureCounter();
     }
 
     /// <summary>
@@ -92,6 +101,19 @@ public sealed class CentralAuditWriter : ICentralAuditWriter
         catch (Exception ex)
         {
             // Audit failure NEVER aborts the user-facing action — swallow and log.
+            // M6 Bundle E (T8): also surface the failure on the central health
+            // counter so a sustained audit-write outage is visible on the
+            // health dashboard rather than disappearing into the log file.
+            try
+            {
+                _failureCounter.Increment();
+            }
+            catch
+            {
+                // Counter must NEVER throw — defence in depth. Even if a
+                // misbehaving custom counter does, swallowing here keeps the
+                // best-effort contract intact.
+            }
             _logger.LogWarning(
                 ex,
                 "CentralAuditWriter failed for EventId {EventId} (Kind={Kind}, Status={Status})",

src/ScadaLink.AuditLog/Central/IAuditCentralHealthSnapshot.cs

@@ -0,0 +1,62 @@
+using ScadaLink.AuditLog.Payload;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E read-side surface exposing the central-side
+/// audit-health counters: <see cref="CentralAuditWriteFailures"/> (every
+/// repository insert throw from <see cref="CentralAuditWriter"/> /
+/// <see cref="AuditLogIngestActor"/>), <see cref="AuditRedactionFailure"/>
+/// (every payload-filter redactor throw on the central path), and
+/// <see cref="SiteAuditTelemetryStalled"/> (per-site latched state from the
+/// <see cref="SiteAuditTelemetryStalledTracker"/>).
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Read-only contract.</b> Implementations expose a point-in-time snapshot
+/// — increments and tracker updates happen through the dedicated counter /
+/// tracker interfaces, not through this surface. Consumers (M7+ central
+/// health pages) read these properties; they never mutate.
+/// </para>
+/// <para>
+/// <b>Why a parallel surface from <see cref="ICentralHealthAggregator"/>.</b>
+/// <see cref="ICentralHealthAggregator"/> aggregates per-site
+/// <c>SiteHealthState</c> reports the SITE emits. The central audit-write
+/// failure / redaction-failure counters originate ON central (no site report
+/// carries them), so they live on a dedicated snapshot rather than being
+/// retro-fitted into a per-site state. The two surfaces will be composed at
+/// the M7 dashboard layer.
+/// </para>
+/// </remarks>
+public interface IAuditCentralHealthSnapshot
+{
+    /// <summary>
+    /// Count of central-side audit-write failures since process start.
+    /// Incremented by every <see cref="CentralAuditWriter"/> /
+    /// <see cref="AuditLogIngestActor"/> repository insert that throws.
+    /// </summary>
+    int CentralAuditWriteFailures { get; }
+
+    /// <summary>
+    /// Count of central-side payload-filter redactor over-redactions since
+    /// process start. Incremented by every header / body / SQL-parameter
+    /// redactor stage that throws (the filter falls back to the
+    /// <c>&lt;redacted: redactor error&gt;</c> marker and never aborts the
+    /// user-facing action). Sites have their own counter
+    /// (<see cref="IAuditRedactionFailureCounter"/>-backed
+    /// <c>SiteHealthReport.AuditRedactionFailure</c>) and the central
+    /// composition root's binding routes ALL central redactor throws
+    /// (CentralAuditWriter + AuditLogIngestActor paths) into this counter.
+    /// </summary>
+    int AuditRedactionFailure { get; }
+
+    /// <summary>
+    /// Per-site latched stalled state: <c>true</c> when the
+    /// <see cref="SiteAuditReconciliationActor"/> has observed two
+    /// consecutive non-draining cycles for that site, <c>false</c> after the
+    /// first draining cycle. Sites absent from the map are interpreted as
+    /// healthy (<c>Stalled=false</c> default). Snapshot is a defensive
+    /// copy — readers must not mutate.
+    /// </summary>
+    IReadOnlyDictionary<string, bool> SiteAuditTelemetryStalled { get; }
+}

src/ScadaLink.AuditLog/Central/ICentralAuditWriteFailureCounter.cs

@@ -0,0 +1,23 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T8) counter sink invoked by central-side audit
+/// writers (<see cref="CentralAuditWriter"/>, <see cref="AuditLogIngestActor"/>)
+/// every time a repository <c>InsertIfNotExistsAsync</c> throws. Mirrors the
+/// site-side <see cref="ScadaLink.AuditLog.Site.IAuditWriteFailureCounter"/>
+/// shape one-for-one — same one-method contract, same NoOp default, same
+/// must-never-abort-the-user-facing-action invariant.
+/// </summary>
+/// <remarks>
+/// Audit-write failures NEVER abort the user-facing action (alog.md §13) —
+/// the writer swallows the exception and surfaces the failure via this counter
+/// instead. A NoOp default is the correct safe fallback while the central
+/// health surface is being wired in; <see cref="AuditCentralHealthSnapshot"/>
+/// is the production binding that routes increments into the aggregated
+/// central health snapshot consumed by future M7+ pages.
+/// </remarks>
+public interface ICentralAuditWriteFailureCounter
+{
+    /// <summary>Increment the central audit-write failure counter by one.</summary>
+    void Increment();
+}

src/ScadaLink.AuditLog/Central/IPullAuditEventsClient.cs

@@ -0,0 +1,45 @@
+using ScadaLink.Commons.Messages.Integration;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Mockable abstraction over the central-side <c>PullAuditEvents</c> gRPC
+/// client surface that <see cref="SiteAuditReconciliationActor"/> uses to
+/// fetch the next reconciliation batch from a specific site. Extracted so the
+/// actor can be unit-tested against an in-memory stub without standing up a
+/// real <c>GrpcChannel</c> per site.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The production implementation (host wiring task) wraps the auto-generated
+/// <c>SiteStreamService.SiteStreamServiceClient</c>, multiplexing one
+/// <c>GrpcChannel</c> per site keyed on
+/// <see cref="SiteEntry.GrpcEndpoint"/>. Until that wiring lands the DI
+/// composition root binds a NoOp default that returns an empty response — the
+/// reconciliation tick is still scheduled and the cursor logic still runs, so
+/// regressions in the actor itself are caught even before the real client
+/// arrives.
+/// </para>
+/// <para>
+/// Implementations MUST NOT throw on transport faults that the actor can
+/// tolerate (connection refused, deadline exceeded). The actor's contract is
+/// "one site's failure doesn't sink the rest of the tick"; an exception still
+/// won't crash the actor (the per-site try/catch catches it), but returning
+/// an empty response on a known-recoverable error keeps the logs cleaner.
+/// </para>
+/// </remarks>
+public interface IPullAuditEventsClient
+{
+    /// <summary>
+    /// Issues a <c>PullAuditEvents</c> RPC against the site whose endpoint
+    /// is registered against <paramref name="siteId"/>. Returns the next
+    /// batch of <see cref="ScadaLink.Commons.Entities.Audit.AuditEvent"/>
+    /// rows ordered oldest-first AND a <c>MoreAvailable</c> flag the actor
+    /// uses to decide whether to fire another pull immediately.
+    /// </summary>
+    Task<PullAuditEventsResponse> PullAsync(
+        string siteId,
+        DateTime sinceUtc,
+        int batchSize,
+        CancellationToken ct);
+}

src/ScadaLink.AuditLog/Central/ISiteEnumerator.cs

@@ -0,0 +1,34 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Enumeration surface consumed by <see cref="SiteAuditReconciliationActor"/> to
+/// discover which sites to poll on each reconciliation tick. Extracted so the
+/// actor can be unit-tested against a static list without depending on the
+/// production <c>ISiteRepository</c> + EF Core DbContext.
+/// </summary>
+/// <remarks>
+/// The production implementation wraps <c>ISiteRepository.GetAllSitesAsync</c>
+/// and projects each <c>Site</c> to a <see cref="SiteEntry"/> using the
+/// site's configured <c>GrpcNodeAAddress</c> (falling back to
+/// <c>GrpcNodeBAddress</c> when NodeA is unset). Sites with NO gRPC address
+/// configured are silently skipped — the reconciliation pull cannot reach
+/// them, but absence of an address is a configuration decision, not a runtime
+/// error.
+/// </remarks>
+public interface ISiteEnumerator
+{
+    /// <summary>
+    /// Returns the current set of sites the reconciliation puller should visit
+    /// on the next tick. Implementations should reflect adds/removes promptly
+    /// — the actor calls this once per tick.
+    /// </summary>
+    Task<IReadOnlyList<SiteEntry>> EnumerateAsync(CancellationToken ct = default);
+}
+
+/// <summary>
+/// One reconciliation target: the site identifier the actor uses as the
+/// cursor key and the gRPC endpoint <see cref="IPullAuditEventsClient"/> dials
+/// to issue the pull. Endpoint is the bare authority (e.g. <c>http://siteA:8083</c>);
+/// transport selection (TLS, keepalive, etc.) is the client's concern.
+/// </summary>
+public sealed record SiteEntry(string SiteId, string GrpcEndpoint);

src/ScadaLink.AuditLog/Central/NoOpCentralAuditWriteFailureCounter.cs

@@ -0,0 +1,17 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Default <see cref="ICentralAuditWriteFailureCounter"/> binding used when
+/// the central health surface (<see cref="AuditCentralHealthSnapshot"/>) has
+/// not been wired (test composition roots, site-only hosts that incidentally
+/// resolve a <see cref="CentralAuditWriter"/>). Drops every increment on the
+/// floor. Mirrors <see cref="ScadaLink.AuditLog.Site.NoOpAuditWriteFailureCounter"/>.
+/// </summary>
+public sealed class NoOpCentralAuditWriteFailureCounter : ICentralAuditWriteFailureCounter
+{
+    /// <inheritdoc/>
+    public void Increment()
+    {
+        // intentional no-op
+    }
+}

src/ScadaLink.AuditLog/Central/SiteAuditReconciliationActor.cs

@@ -0,0 +1,332 @@
+using Akka.Actor;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Options;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Central singleton (M6 Bundle B) that drives the audit-log reconciliation
+/// pull loop. On a configurable timer (default 5 minutes) the actor walks every
+/// known site, asks the site for any <see cref="AuditEvent"/> rows with
+/// <see cref="AuditEvent.OccurredAtUtc"/> &gt;= the site's last reconciled
+/// cursor, ingests them idempotently into the central
+/// <see cref="IAuditLogRepository"/>, and advances the cursor.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Self-healing telemetry, not a dispatcher.</b> The push path
+/// (<see cref="ScadaLink.AuditLog.Site.Telemetry.SiteAuditTelemetryActor"/> +
+/// <c>IngestAuditEvents</c>) is the primary mechanism. This actor exists so a
+/// missed push (gRPC blip, central restart, site offline) is eventually
+/// repaired by central re-pulling whatever the site still has in
+/// <c>Pending</c>/<c>Forwarded</c> state. Idempotency on
+/// <see cref="AuditEvent.EventId"/> (M2 Bundle A's race-fix) makes duplicate
+/// arrivals from both paths a silent no-op.
+/// </para>
+/// <para>
+/// <b>Cursor lifetime.</b> The per-site <c>LastReconciledAt</c> watermark is
+/// kept in-memory for the actor's lifetime. The cluster singleton normally
+/// survives the host process; on a deliberate failover OR a singleton restart
+/// the cursors reset to <see cref="DateTime.MinValue"/>. That is conservative
+/// but correct — the next tick simply asks for everything the site still has,
+/// and idempotent ingest swallows the dupes. Persisting cursors to MS SQL was
+/// considered and rejected for M6: the cost of a write per tick outweighs the
+/// rare benefit of avoiding one over-broad pull after a restart.
+/// </para>
+/// <para>
+/// <b>Stalled detection.</b> The brief calls a site "stalled" when two
+/// consecutive pull cycles BOTH return non-empty AND <c>MoreAvailable=true</c>
+/// — i.e. the backlog isn't draining. The actor publishes
+/// <see cref="SiteAuditTelemetryStalledChanged"/> on the actor system's
+/// EventStream so a future <c>ICentralHealthCollector</c> bridge (M6 Bundle E)
+/// can flip the health metric without coupling this actor to the health
+/// collection surface today.
+/// </para>
+/// <para>
+/// <b>Failure isolation.</b> A single site that throws (DNS, transport,
+/// repository write) must NOT prevent other sites from being polled on the
+/// same tick. The per-site work runs inside its own try/catch; the actor's
+/// supervisor strategy keeps it alive across any leaked exception with
+/// <see cref="Akka.Actor.SupervisorStrategy.DefaultDecider"/>'s Restart
+/// semantics — restart resets the in-memory cursors, but as noted above that's
+/// a safe (over-pull, idempotent) recovery.
+/// </para>
+/// <para>
+/// <b>DI scopes.</b> <see cref="IAuditLogRepository"/> is a scoped EF Core
+/// service registered by <c>AddConfigurationDatabase</c>. The singleton actor
+/// opens one DI scope per tick and reuses the same repository across all
+/// sites in that tick — one DbContext per tick mirrors the
+/// <c>AuditLogIngestActor</c> + <c>NotificationOutboxActor</c> pattern.
+/// </para>
+/// </remarks>
+public class SiteAuditReconciliationActor : ReceiveActor
+{
+    private readonly ISiteEnumerator _sites;
+    private readonly IPullAuditEventsClient _client;
+    private readonly IServiceProvider _services;
+    private readonly SiteAuditReconciliationOptions _options;
+    private readonly ILogger<SiteAuditReconciliationActor> _logger;
+
+    /// <summary>
+    /// Per-site reconciliation watermark — the highest
+    /// <see cref="AuditEvent.OccurredAtUtc"/> seen for that site on a previous
+    /// tick. Asking for <c>OccurredAtUtc &gt;= cursor</c> rather than &gt;
+    /// is the site contract (<see cref="ScadaLink.Commons.Interfaces.Services.ISiteAuditQueue.ReadPendingSinceAsync"/>);
+    /// duplicate-with-same-timestamp rows are filtered out by the idempotent
+    /// repository write.
+    /// </summary>
+    private readonly Dictionary<string, DateTime> _cursors = new();
+
+    /// <summary>
+    /// Per-site count of consecutive non-draining cycles. Resets to zero on the
+    /// first draining (or empty) cycle.
+    /// </summary>
+    private readonly Dictionary<string, int> _nonDrainingCycles = new();
+
+    /// <summary>
+    /// Per-site latched stalled state — used so the actor only publishes a
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> transition when the
+    /// stalled flag actually changes, not on every tick while stalled.
+    /// </summary>
+    private readonly Dictionary<string, bool> _stalled = new();
+
+    private ICancelable? _timer;
+
+    public SiteAuditReconciliationActor(
+        ISiteEnumerator sites,
+        IPullAuditEventsClient client,
+        IServiceProvider services,
+        IOptions<SiteAuditReconciliationOptions> options,
+        ILogger<SiteAuditReconciliationActor> logger)
+    {
+        ArgumentNullException.ThrowIfNull(sites);
+        ArgumentNullException.ThrowIfNull(client);
+        ArgumentNullException.ThrowIfNull(services);
+        ArgumentNullException.ThrowIfNull(options);
+        ArgumentNullException.ThrowIfNull(logger);
+
+        _sites = sites;
+        _client = client;
+        _services = services;
+        _options = options.Value;
+        _logger = logger;
+
+        ReceiveAsync<ReconciliationTick>(_ => OnTickAsync());
+    }
+
+    protected override void PreStart()
+    {
+        base.PreStart();
+        var interval = _options.ReconciliationInterval;
+        _timer = Context.System.Scheduler.ScheduleTellRepeatedlyCancelable(
+            initialDelay: interval,
+            interval: interval,
+            receiver: Self,
+            message: ReconciliationTick.Instance,
+            sender: Self);
+    }
+
+    protected override void PostStop()
+    {
+        _timer?.Cancel();
+        base.PostStop();
+    }
+
+    private async Task OnTickAsync()
+    {
+        // Capture EventStream BEFORE the first await. Accessing Context (and
+        // therefore Context.System) after an await is unsafe because Akka's
+        // ActorBase.Context throws "no active ActorContext" once the
+        // continuation runs on a thread that isn't currently dispatching this
+        // actor — mirrors the AuditLogPurgeActor.OnTickAsync fix and the
+        // AuditLogIngestActor.OnIngestAsync Sender-capture pattern.
+        var eventStream = Context.System.EventStream;
+
+        IReadOnlyList<SiteEntry> sites;
+        try
+        {
+            sites = await _sites.EnumerateAsync().ConfigureAwait(false);
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex, "Site enumeration failed; skipping reconciliation tick.");
+            return;
+        }
+
+        if (sites.Count == 0)
+        {
+            return;
+        }
+
+        IServiceScope? scope = null;
+        IAuditLogRepository repository;
+        try
+        {
+            scope = _services.CreateScope();
+            repository = scope.ServiceProvider.GetRequiredService<IAuditLogRepository>();
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex, "Failed to resolve IAuditLogRepository for reconciliation tick.");
+            scope?.Dispose();
+            return;
+        }
+
+        try
+        {
+            foreach (var site in sites)
+            {
+                try
+                {
+                    await PullSiteAsync(site, repository, eventStream).ConfigureAwait(false);
+                }
+                catch (Exception ex)
+                {
+                    // Catch-all per the failure-isolation invariant: one site's
+                    // fault must not sink the rest of the tick. The cursor for
+                    // the failing site is left at its previous value so the
+                    // next tick retries the same window.
+                    _logger.LogWarning(
+                        ex,
+                        "Reconciliation pull failed for site {SiteId}; other sites continue.",
+                        site.SiteId);
+                }
+            }
+        }
+        finally
+        {
+            scope.Dispose();
+        }
+    }
+
+    /// <summary>
+    /// Issues one <c>PullAuditEvents</c> RPC against the site, ingests the
+    /// returned rows idempotently into the central repository, and advances
+    /// the cursor based on the maximum <see cref="AuditEvent.OccurredAtUtc"/>
+    /// observed. The brief's "saturate until backlog clears" intent is met by
+    /// the natural cadence — each tick issues one pull, and a backed-up site
+    /// drains across consecutive ticks. The stalled signal (two non-draining
+    /// ticks in a row) surfaces when that drain isn't keeping up.
+    /// </summary>
+    private async Task PullSiteAsync(SiteEntry site, IAuditLogRepository repository, Akka.Event.EventStream eventStream)
+    {
+        var since = _cursors.TryGetValue(site.SiteId, out var c) ? c : DateTime.MinValue;
+        var response = await _client.PullAsync(
+            site.SiteId, since, _options.BatchSize, CancellationToken.None)
+            .ConfigureAwait(false);
+
+        var maxOccurred = since;
+        var nowUtc = DateTime.UtcNow;
+        foreach (var evt in response.Events)
+        {
+            try
+            {
+                // Idempotent repository write: duplicate EventIds (from a
+                // concurrent push, or a retry of this very pull) collapse to
+                // a no-op courtesy of M2 Bundle A's race-fix on
+                // InsertIfNotExistsAsync.
+                var ingested = evt with { IngestedAtUtc = nowUtc };
+                await repository.InsertIfNotExistsAsync(ingested).ConfigureAwait(false);
+            }
+            catch (Exception ex)
+            {
+                // Per-row catch so one bad event does not abandon the rest of
+                // the batch. The cursor still advances based on OccurredAtUtc
+                // — the row was returned by the site, so the next tick won't
+                // re-fetch it; if it permanently fails to persist, that's an
+                // operational concern surfaced by the log, not a hot-loop
+                // trigger.
+                _logger.LogError(
+                    ex,
+                    "Reconciliation ingest failed for AuditEvent {EventId} from site {SiteId}.",
+                    evt.EventId,
+                    site.SiteId);
+            }
+
+            if (evt.OccurredAtUtc > maxOccurred)
+            {
+                maxOccurred = evt.OccurredAtUtc;
+            }
+        }
+
+        _cursors[site.SiteId] = maxOccurred;
+
+        var nonDraining = response.MoreAvailable && response.Events.Count > 0;
+        UpdateStalledState(site.SiteId, draining: !nonDraining, eventStream);
+    }
+
+    /// <summary>
+    /// Flips the per-site stalled flag based on whether this tick drained the
+    /// queue. A "draining" cycle is one where the server reported no more rows
+    /// available OR returned zero events. A "non-draining" cycle is the
+    /// inverse (events returned AND <c>MoreAvailable=true</c>).
+    /// </summary>
+    /// <remarks>
+    /// The state machine: counter increments on each consecutive non-draining
+    /// tick. On reaching <see cref="SiteAuditReconciliationOptions.StalledAfterNonDrainingCycles"/>
+    /// the actor latches <c>Stalled=true</c> and publishes the transition; on
+    /// any subsequent draining tick the counter resets to zero AND, if the
+    /// latch is currently true, the actor publishes <c>Stalled=false</c>. Only
+    /// transitions are published — repeated ticks in the same state are
+    /// silent so a downstream subscriber doesn't see a flood of redundant
+    /// notifications.
+    /// </remarks>
+    private void UpdateStalledState(string siteId, bool draining, Akka.Event.EventStream eventStream)
+    {
+        var wasStalled = _stalled.TryGetValue(siteId, out var prior) && prior;
+
+        if (draining)
+        {
+            _nonDrainingCycles[siteId] = 0;
+            if (wasStalled)
+            {
+                _stalled[siteId] = false;
+                eventStream.Publish(
+                    new SiteAuditTelemetryStalledChanged(siteId, Stalled: false));
+            }
+            return;
+        }
+
+        var consecutive = _nonDrainingCycles.GetValueOrDefault(siteId) + 1;
+        _nonDrainingCycles[siteId] = consecutive;
+
+        if (consecutive >= _options.StalledAfterNonDrainingCycles && !wasStalled)
+        {
+            _stalled[siteId] = true;
+            eventStream.Publish(
+                new SiteAuditTelemetryStalledChanged(siteId, Stalled: true));
+        }
+    }
+
+    /// <summary>
+    /// Resume on any unhandled exception inside the receive — the singleton
+    /// MUST stay alive even if the per-tick try/catch leaks. Restart would
+    /// reset the cursors (safe but wasteful); Resume preserves them.
+    /// </summary>
+    protected override SupervisorStrategy SupervisorStrategy()
+    {
+        return new OneForOneStrategy(
+            maxNrOfRetries: 0,
+            withinTimeRange: TimeSpan.Zero,
+            decider: Akka.Actor.SupervisorStrategy.DefaultDecider);
+    }
+
+    /// <summary>Self-tick triggering a reconciliation pass across all sites.</summary>
+    internal sealed class ReconciliationTick
+    {
+        public static readonly ReconciliationTick Instance = new();
+        private ReconciliationTick() { }
+    }
+}
+
+/// <summary>
+/// Published on the actor system EventStream when a site's reconciliation
+/// puller transitions into or out of the "stalled" state (backlog not
+/// draining across multiple cycles). The M6 Bundle E central health collector
+/// will subscribe to this and surface
+/// <c>SiteAuditTelemetryStalled</c> on the health-report payload.
+/// </summary>
+public sealed record SiteAuditTelemetryStalledChanged(string SiteId, bool Stalled);

src/ScadaLink.AuditLog/Central/SiteAuditReconciliationOptions.cs

@@ -0,0 +1,60 @@
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Tuning knobs for the central <see cref="SiteAuditReconciliationActor"/> singleton.
+/// Defaults mirror the M6 Bundle B brief: pull every 5 minutes per site, 256 rows per
+/// batch, declare a site "stalled" after two consecutive pull cycles return non-empty
+/// AND <c>MoreAvailable=true</c> (the backlog is not draining).
+/// </summary>
+/// <remarks>
+/// <para>
+/// Per the M6 plan the reconciliation actor is the fallback when push telemetry is
+/// lost; it is intentionally low-frequency. Lowering
+/// <see cref="ReconciliationIntervalSeconds"/> in production trades MS SQL load for
+/// fresher self-healing — keep the default unless a deployment can prove the extra
+/// load is acceptable.
+/// </para>
+/// <para>
+/// <see cref="StalledAfterNonDrainingCycles"/> = 2 because a single non-draining
+/// cycle can happen on a surge (e.g. a backed-up site replays its hot queue); the
+/// stalled signal should only fire when the backlog persists across cycles, which is
+/// the symptom the central health surface is asking us to detect.
+/// </para>
+/// </remarks>
+public sealed class SiteAuditReconciliationOptions
+{
+    /// <summary>
+    /// Period of the reconciliation tick. Each tick visits every known site once.
+    /// </summary>
+    public int ReconciliationIntervalSeconds { get; set; } = 300;
+
+    /// <summary>
+    /// Test-only override for finer control over the tick cadence than
+    /// whole-second resolution allows. When non-null, takes precedence over
+    /// <see cref="ReconciliationIntervalSeconds"/>. Not bound from config —
+    /// production config exposes <see cref="ReconciliationIntervalSeconds"/>
+    /// only.
+    /// </summary>
+    public TimeSpan? ReconciliationIntervalOverride { get; set; }
+
+    /// <summary>
+    /// Resolves the effective tick interval, honouring the test override when
+    /// set. Falls back to <see cref="ReconciliationIntervalSeconds"/>.
+    /// </summary>
+    public TimeSpan ReconciliationInterval =>
+        ReconciliationIntervalOverride ?? TimeSpan.FromSeconds(ReconciliationIntervalSeconds);
+
+    /// <summary>
+    /// Maximum number of <see cref="ScadaLink.Commons.Entities.Audit.AuditEvent"/>
+    /// rows requested in a single <c>PullAuditEvents</c> RPC call.
+    /// </summary>
+    public int BatchSize { get; set; } = 256;
+
+    /// <summary>
+    /// Number of consecutive non-draining cycles (events returned AND
+    /// <c>MoreAvailable=true</c>) that must accumulate for a site before the actor
+    /// publishes <c>SiteAuditTelemetryStalledChanged(Stalled: true)</c> on the
+    /// EventStream.
+    /// </summary>
+    public int StalledAfterNonDrainingCycles { get; set; } = 2;
+}

src/ScadaLink.AuditLog/Central/SiteAuditTelemetryStalledTracker.cs

@@ -0,0 +1,188 @@
+using System.Collections.Concurrent;
+using Akka.Actor;
+using Akka.Event;
+
+namespace ScadaLink.AuditLog.Central;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T7) — central singleton that subscribes to the
+/// actor system's EventStream for <see cref="SiteAuditTelemetryStalledChanged"/>
+/// publications and maintains a per-site latched stalled-state map readable
+/// via <see cref="Snapshot"/>. Consumed by the M6 Bundle E
+/// <see cref="AuditCentralHealthSnapshot"/> aggregator so the central health
+/// surface can surface per-site "reconciliation isn't draining" without
+/// coupling the publisher (<see cref="SiteAuditReconciliationActor"/>) to the
+/// health collection plumbing.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Why an internal actor.</b> Akka.NET's <see cref="EventStream"/> only
+/// supports <see cref="IActorRef"/> subscribers — there is no callback or
+/// channel-based overload. The tracker therefore spawns a small subscriber
+/// actor that forwards each event into the shared
+/// <see cref="ConcurrentDictionary{TKey,TValue}"/> on the actor's thread, and
+/// readers (<see cref="Snapshot"/>) take a copy off that dictionary on any
+/// thread. Mirrors the <c>DeadLetterMonitorActor</c> shape — subscribe in
+/// <see cref="ActorBase.PreStart"/>, unsubscribe in
+/// <see cref="ActorBase.PostStop"/>, which the tracker triggers via a Stop
+/// at <see cref="Dispose"/>.
+/// </para>
+/// <para>
+/// <b>Per-site latching.</b> The publisher (<see cref="SiteAuditReconciliationActor"/>)
+/// only publishes on stalled-state transitions, so the dictionary is the
+/// authoritative latched state. Sites that have never published are absent
+/// from the snapshot — the consumer surface treats absence as
+/// <c>Stalled=false</c> (default healthy), the same default the reconciliation
+/// actor's own internal latch uses.
+/// </para>
+/// <para>
+/// <b>Singleton lifecycle.</b> Registered as a singleton via
+/// <see cref="ServiceCollectionExtensions.AddAuditLogCentralMaintenance"/>;
+/// <see cref="Dispose"/> tears the internal subscriber down at host shutdown.
+/// </para>
+/// </remarks>
+public sealed class SiteAuditTelemetryStalledTracker : IDisposable
+{
+    private readonly EventStream _eventStream;
+    private readonly ConcurrentDictionary<string, bool> _state = new();
+    private readonly IActorRef? _subscriber;
+    private readonly AuditCentralHealthSnapshot? _snapshot;
+    private bool _disposed;
+
+    /// <summary>
+    /// Construct around a bare <see cref="EventStream"/>. Intended for unit
+    /// tests where the caller wants to publish events without standing up an
+    /// actor system — the tracker registers a transient subscriber actor only
+    /// if the supplied stream is backed by an actor system. In the bare-stream
+    /// mode (no actor system) the tracker still exposes the
+    /// <see cref="Snapshot"/> surface but cannot self-subscribe; production
+    /// callers always go through <see cref="SiteAuditTelemetryStalledTracker(ActorSystem)"/>.
+    /// </summary>
+    /// <remarks>
+    /// Subscribing to <see cref="EventStream"/> requires an <see cref="IActorRef"/>,
+    /// which can only be created from an <see cref="ActorSystem"/>. The bare-
+    /// stream ctor therefore can NOT itself wire the subscriber — tests that
+    /// want event-driven updates must use the ActorSystem ctor (or push state
+    /// directly via <see cref="Apply"/>). The tests in
+    /// <c>SiteAuditTelemetryStalledTrackerTests</c> use the ActorSystem ctor
+    /// via Akka.TestKit so they exercise the production subscribe path.
+    /// </remarks>
+    public SiteAuditTelemetryStalledTracker(EventStream eventStream)
+        : this(eventStream, snapshot: null)
+    {
+    }
+
+    /// <summary>
+    /// Bare-stream ctor with an optional snapshot sink — the central
+    /// composition root passes the singleton
+    /// <see cref="AuditCentralHealthSnapshot"/> so every dictionary update
+    /// also lands on the central health surface. The bare ctor still cannot
+    /// subscribe (no actor system), but tests that drive the tracker via
+    /// <see cref="Apply"/> get the snapshot push for free.
+    /// </summary>
+    public SiteAuditTelemetryStalledTracker(EventStream eventStream, AuditCentralHealthSnapshot? snapshot)
+    {
+        _eventStream = eventStream ?? throw new ArgumentNullException(nameof(eventStream));
+        // No subscriber actor — see the remarks on the parameterless overload.
+        _subscriber = null;
+        _snapshot = snapshot;
+    }
+
+    /// <summary>
+    /// Production ctor: subscribes a small internal actor to the supplied
+    /// system's EventStream so every published
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> updates the latched
+    /// per-site map. <see cref="Dispose"/> tears the subscriber down.
+    /// </summary>
+    public SiteAuditTelemetryStalledTracker(ActorSystem actorSystem)
+        : this(actorSystem, snapshot: null)
+    {
+    }
+
+    /// <summary>
+    /// Production ctor with a snapshot sink — every observed
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> is mirrored onto the
+    /// shared <see cref="AuditCentralHealthSnapshot"/> so the central health
+    /// surface sees per-site stalled state without re-reading the tracker.
+    /// </summary>
+    public SiteAuditTelemetryStalledTracker(ActorSystem actorSystem, AuditCentralHealthSnapshot? snapshot)
+    {
+        ArgumentNullException.ThrowIfNull(actorSystem);
+        _eventStream = actorSystem.EventStream;
+        _snapshot = snapshot;
+        // Anonymous subscriber actor scoped to the system; props build it
+        // with a callback into THIS tracker's Apply method so the actor's
+        // single-threaded receive serialises every dictionary write.
+        _subscriber = actorSystem.ActorOf(
+            Props.Create(() => new StalledChangedSubscriber(this)),
+            name: $"site-audit-stalled-tracker-{Guid.NewGuid():N}");
+        // Subscribe synchronously from the ctor so the subscription is in
+        // place before the tracker is returned to the caller — the actor's
+        // own PreStart runs asynchronously and would otherwise race the
+        // first publish. EventStream.Subscribe is thread-safe.
+        _eventStream.Subscribe(_subscriber, typeof(SiteAuditTelemetryStalledChanged));
+    }
+
+    /// <summary>
+    /// Returns a defensive copy of the per-site latched stalled state.
+    /// Absent sites are interpreted as <c>Stalled=false</c> by consumers.
+    /// </summary>
+    public IReadOnlyDictionary<string, bool> Snapshot() =>
+        new Dictionary<string, bool>(_state);
+
+    /// <summary>
+    /// Applied by the internal subscriber actor on every
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> publication. Exposed
+    /// internally so tests against the bare-stream ctor can still drive the
+    /// tracker, but the production path always goes through the actor.
+    /// </summary>
+    internal void Apply(SiteAuditTelemetryStalledChanged evt)
+    {
+        if (evt is null) return;
+        _state[evt.SiteId] = evt.Stalled;
+        // Mirror into the central health snapshot if wired so a reader of
+        // IAuditCentralHealthSnapshot sees the same per-site state without
+        // a second lookup. Snapshot is optional (test composition roots may
+        // skip it) so the null-coalesce is the safe path.
+        _snapshot?.ApplyStalled(evt);
+    }
+
+    public void Dispose()
+    {
+        if (_disposed) return;
+        _disposed = true;
+        if (_subscriber is not null)
+        {
+            // Unsubscribe runs in PostStop on the subscriber actor; Stop is
+            // fire-and-forget but the actor's PostStop hook is guaranteed to
+            // run before its mailbox is collected.
+            _subscriber.Tell(PoisonPill.Instance);
+        }
+    }
+
+    /// <summary>
+    /// Internal subscriber actor — receives every
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> off the EventStream and
+    /// forwards it into the parent <see cref="SiteAuditTelemetryStalledTracker"/>.
+    /// Unlike <c>DeadLetterMonitorActor</c>, the subscription is registered by
+    /// the tracker constructor BEFORE this actor begins processing messages so
+    /// publishes that arrive between actor creation and PreStart cannot be
+    /// missed. Unsubscribe still runs in <see cref="PostStop"/>.
+    /// </summary>
+    private sealed class StalledChangedSubscriber : ReceiveActor
+    {
+        private readonly SiteAuditTelemetryStalledTracker _parent;
+
+        public StalledChangedSubscriber(SiteAuditTelemetryStalledTracker parent)
+        {
+            _parent = parent;
+            Receive<SiteAuditTelemetryStalledChanged>(evt => _parent.Apply(evt));
+        }
+
+        protected override void PostStop()
+        {
+            Context.System.EventStream.Unsubscribe(Self, typeof(SiteAuditTelemetryStalledChanged));
+            base.PostStop();
+        }
+    }
+}

src/ScadaLink.AuditLog/ServiceCollectionExtensions.cs

@@ -1,6 +1,7 @@
 using Microsoft.Extensions.Configuration;
 using Microsoft.Extensions.DependencyInjection;
 using Microsoft.Extensions.DependencyInjection.Extensions;
+using Microsoft.Extensions.Hosting;
 using Microsoft.Extensions.Logging;
 using Microsoft.Extensions.Options;
 using ScadaLink.AuditLog.Central;
@@ -43,6 +44,9 @@ public static class ServiceCollectionExtensions
     /// <summary>Configuration section bound to <see cref="SiteAuditTelemetryOptions"/>.</summary>
     public const string SiteTelemetrySectionName = "AuditLog:SiteTelemetry";
 
+    /// <summary>Configuration section bound to <see cref="AuditLogPartitionMaintenanceOptions"/>.</summary>
+    public const string PartitionMaintenanceSectionName = "AuditLog:PartitionMaintenance";
+
     /// <summary>
     /// Registers the Audit Log (#23) component services: options, the site
     /// SQLite writer chain (primary + ring fallback + failure-counter sink),
@@ -151,6 +155,13 @@ public static class ServiceCollectionExtensions
         services.AddSingleton<ICachedCallLifecycleObserver>(
             sp => sp.GetRequiredService<CachedCallLifecycleBridge>());
 
+        // M6 Bundle E (T8): central audit-write failure counter — NoOp default
+        // for site/test composition roots that don't wire the central health
+        // snapshot. AddAuditLogCentralMaintenance below replaces this binding
+        // with the AuditCentralHealthSnapshot implementation so increments
+        // surface on the central dashboard.
+        services.TryAddSingleton<ICentralAuditWriteFailureCounter, NoOpCentralAuditWriteFailureCounter>();
+
         // M4 Bundle B: central direct-write audit writer used by
         // NotificationOutboxActor (Bundle B) and Inbound API (Bundle C/D) to
         // emit AuditLog rows that originate ON central, not via site telemetry.
@@ -163,10 +174,13 @@ public static class ServiceCollectionExtensions
         // Bundle C (M5-T6): wire the IAuditPayloadFilter into the factory so
         // NotificationOutboxActor + Inbound API rows are truncated + redacted
         // before they hit MS SQL.
+        // M6 Bundle E (T8): also wire the ICentralAuditWriteFailureCounter
+        // so swallowed repo throws bump the central health counter.
         services.AddSingleton<ICentralAuditWriter>(sp => new CentralAuditWriter(
             sp,
             sp.GetRequiredService<ILogger<CentralAuditWriter>>(),
-            sp.GetRequiredService<IAuditPayloadFilter>()));
+            sp.GetRequiredService<IAuditPayloadFilter>(),
+            sp.GetRequiredService<ICentralAuditWriteFailureCounter>()));
 
         return services;
     }
@@ -214,6 +228,80 @@ public static class ServiceCollectionExtensions
             ServiceDescriptor.Singleton<IAuditWriteFailureCounter, HealthMetricsAuditWriteFailureCounter>());
         services.Replace(
             ServiceDescriptor.Singleton<IAuditRedactionFailureCounter, HealthMetricsAuditRedactionFailureCounter>());
+        // M6 Bundle E (T6): the site-side backlog reporter polls the
+        // SqliteAuditWriter every 30 s and pushes the snapshot into the
+        // collector so the next SiteHealthReport carries a fresh
+        // SiteAuditBacklog field. Registered alongside the other site-only
+        // metric bridges so AddAuditLog (which runs on central too) stays
+        // free of hosted-service registrations that would resolve a missing
+        // ISiteHealthCollector on central.
+        services.AddHostedService<SiteAuditBacklogReporter>();
+        return services;
+    }
+
+    /// <summary>
+    /// Audit Log (#23) M6-T5 Bundle D — central-only registration for the
+    /// <see cref="AuditLogPartitionMaintenanceService"/> hosted service plus
+    /// its <see cref="AuditLogPartitionMaintenanceOptions"/> binding. Must be
+    /// called from the Central role's composition root (not from a site
+    /// composition root); the underlying <c>IPartitionMaintenance</c>
+    /// implementation is registered by <c>AddConfigurationDatabase</c> and
+    /// only exists on the central node.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// Separated from <see cref="AddAuditLog"/> because <c>AddAuditLog</c> is
+    /// also invoked from site composition roots — silently starting a
+    /// hosted service that resolves an unregistered dependency on a site
+    /// would fail every tick. Keeping the central-only registration in its
+    /// own helper preserves the "every <c>Add*</c> call is safe to issue
+    /// from any composition root" invariant.
+    /// </para>
+    /// </remarks>
+    public static IServiceCollection AddAuditLogCentralMaintenance(
+        this IServiceCollection services,
+        IConfiguration config)
+    {
+        ArgumentNullException.ThrowIfNull(services);
+        ArgumentNullException.ThrowIfNull(config);
+
+        services.AddOptions<AuditLogPartitionMaintenanceOptions>()
+            .Bind(config.GetSection(PartitionMaintenanceSectionName));
+        services.AddHostedService<AuditLogPartitionMaintenanceService>();
+
+        // M6 Bundle E (T8 + T9): central health snapshot — a single object
+        // that owns the CentralAuditWriteFailures + AuditRedactionFailure
+        // Interlocked counters AND surfaces them on
+        // IAuditCentralHealthSnapshot. The same instance is bound to BOTH
+        // writer-side interfaces (ICentralAuditWriteFailureCounter +
+        // IAuditRedactionFailureCounter) so every central-side increment
+        // routes into the shared counters; site nodes keep their existing
+        // Site bridges (registered by AddAuditLogHealthMetricsBridge) so
+        // the same counter type does not shadow the site-side metric.
+        // The snapshot itself has no actor-system dependency — the
+        // per-site stalled latch is fed by SiteAuditTelemetryStalledTracker
+        // which the Akka bootstrap wires up after ActorSystem.Create returns
+        // (the tracker is NOT registered here because its construction
+        // requires ActorSystem, which is not a DI-resolvable singleton).
+        services.AddSingleton<AuditCentralHealthSnapshot>();
+        services.AddSingleton<IAuditCentralHealthSnapshot>(
+            sp => sp.GetRequiredService<AuditCentralHealthSnapshot>());
+        services.Replace(ServiceDescriptor.Singleton<ICentralAuditWriteFailureCounter>(
+            sp => sp.GetRequiredService<AuditCentralHealthSnapshot>()));
+        // M6 Bundle E (T9): override the NoOp IAuditRedactionFailureCounter
+        // (registered by AddAuditLog) with the CentralAuditRedactionFailureCounter
+        // bridge so payload-filter throws on CentralAuditWriter /
+        // AuditLogIngestActor paths surface on the central dashboard. The
+        // bridge is a thin wrapper around the AuditCentralHealthSnapshot
+        // singleton so all central redactor failures route into the same
+        // counter as CentralAuditWriteFailures. The site composition root
+        // overrides this binding AGAIN via AddAuditLogHealthMetricsBridge —
+        // central nodes do not call that bridge, so this is the final
+        // binding on a central host. Mirrors the M5 Bundle C
+        // HealthMetricsAuditRedactionFailureCounter shape one-for-one.
+        services.Replace(ServiceDescriptor.Singleton<IAuditRedactionFailureCounter,
+            CentralAuditRedactionFailureCounter>());
+
         return services;
     }
 }

src/ScadaLink.AuditLog/Site/SiteAuditBacklogReporter.cs

@@ -0,0 +1,133 @@
+using Microsoft.Extensions.Hosting;
+using Microsoft.Extensions.Logging;
+using ScadaLink.Commons.Interfaces.Services;
+using ScadaLink.HealthMonitoring;
+
+namespace ScadaLink.AuditLog.Site;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T6) — site-side hosted service that
+/// periodically pulls a backlog snapshot from <see cref="ISiteAuditQueue"/>
+/// and pushes it into <see cref="ISiteHealthCollector"/> so the next
+/// <see cref="ISiteHealthCollector.CollectReport"/> emits a fresh
+/// <c>SiteAuditBacklog</c> field on the site health report.
+/// </summary>
+/// <remarks>
+/// <para>
+/// <b>Why a hosted service, not the report sender.</b> Querying SQLite for the
+/// backlog requires the queue's write lock; doing it inline in
+/// <see cref="ISiteHealthCollector.CollectReport"/> would couple the collector
+/// to <see cref="ISiteAuditQueue"/> and turn an in-memory snapshot read into
+/// a synchronous I/O call on the report path. The hosted-service pattern keeps
+/// the report path pure and the SQL probe off the report timing budget.
+/// </para>
+/// <para>
+/// <b>Cadence.</b> 30 s by default — coarse enough to amortise the SQL probe
+/// across many reports, fine enough that the central dashboard never lags by
+/// more than one health-report interval. Tunable via
+/// <see cref="ScadaLink.AuditLog.Site.SqliteAuditWriterOptions"/> in a follow-up
+/// if ops needs a different cadence; for M6 we hard-code the value because the
+/// brief calls it out explicitly.
+/// </para>
+/// <para>
+/// <b>Failure containment.</b> The probe call is wrapped in a try/catch so a
+/// transient SQLite error never tears down the hosted service — the next tick
+/// retries. Mirrors <see cref="ScadaLink.AuditLog.Central.AuditLogPartitionMaintenanceService"/>'s
+/// "exception logged, not propagated" contract.
+/// </para>
+/// </remarks>
+public sealed class SiteAuditBacklogReporter : IHostedService, IDisposable
+{
+    /// <summary>
+    /// Default poll cadence. Half a typical 60 s health-report interval keeps
+    /// the snapshot fresh without spinning the SQL probe more often than
+    /// necessary.
+    /// </summary>
+    internal static readonly TimeSpan DefaultRefreshInterval = TimeSpan.FromSeconds(30);
+
+    private readonly ISiteAuditQueue _queue;
+    private readonly ISiteHealthCollector _collector;
+    private readonly ILogger<SiteAuditBacklogReporter> _logger;
+    private readonly TimeSpan _refreshInterval;
+    private CancellationTokenSource? _cts;
+    private Task? _loop;
+
+    public SiteAuditBacklogReporter(
+        ISiteAuditQueue queue,
+        ISiteHealthCollector collector,
+        ILogger<SiteAuditBacklogReporter> logger,
+        TimeSpan? refreshInterval = null)
+    {
+        _queue = queue ?? throw new ArgumentNullException(nameof(queue));
+        _collector = collector ?? throw new ArgumentNullException(nameof(collector));
+        _logger = logger ?? throw new ArgumentNullException(nameof(logger));
+        _refreshInterval = refreshInterval ?? DefaultRefreshInterval;
+    }
+
+    /// <inheritdoc />
+    public Task StartAsync(CancellationToken ct)
+    {
+        // Linked CTS lets StopAsync's cancellation AND the host's shutdown
+        // token both terminate the loop; either side firing aborts the
+        // pending Task.Delay.
+        _cts = CancellationTokenSource.CreateLinkedTokenSource(ct);
+        _loop = Task.Run(() => RunLoopAsync(_cts.Token));
+        return Task.CompletedTask;
+    }
+
+    private async Task RunLoopAsync(CancellationToken ct)
+    {
+        // First tick runs immediately so the very first health report after
+        // process start carries a real backlog snapshot — without this the
+        // dashboard would show null for the first 30 s after a deploy.
+        await SafeProbeAsync(ct).ConfigureAwait(false);
+
+        while (!ct.IsCancellationRequested)
+        {
+            try
+            {
+                await Task.Delay(_refreshInterval, ct).ConfigureAwait(false);
+            }
+            catch (OperationCanceledException)
+            {
+                break;
+            }
+
+            await SafeProbeAsync(ct).ConfigureAwait(false);
+        }
+    }
+
+    private async Task SafeProbeAsync(CancellationToken ct)
+    {
+        try
+        {
+            var snapshot = await _queue.GetBacklogStatsAsync(ct).ConfigureAwait(false);
+            _collector.UpdateSiteAuditBacklog(snapshot);
+        }
+        catch (OperationCanceledException)
+        {
+            // Shutdown — let the outer loop exit cleanly.
+            throw;
+        }
+        catch (Exception ex)
+        {
+            // Catch-all is deliberate: the hosted service must survive every
+            // class of probe failure (transient SQLite lock contention, disk
+            // I/O hiccup, …) so the next tick gets a chance.
+            _logger.LogWarning(ex, "SiteAuditBacklogReporter probe failed; next tick will retry.");
+        }
+    }
+
+    /// <inheritdoc />
+    public Task StopAsync(CancellationToken ct)
+    {
+        _cts?.Cancel();
+        return _loop ?? Task.CompletedTask;
+    }
+
+    /// <inheritdoc />
+    public void Dispose()
+    {
+        _cts?.Dispose();
+    }
+}

src/ScadaLink.AuditLog/Site/SqliteAuditWriter.cs

@@ -2,9 +2,9 @@ using System.Threading.Channels;
 using Microsoft.Data.Sqlite;
 using Microsoft.Extensions.Logging;
 using Microsoft.Extensions.Options;
-using ScadaLink.AuditLog.Site.Telemetry;
 using ScadaLink.Commons.Entities.Audit;
 using ScadaLink.Commons.Interfaces.Services;
+using ScadaLink.Commons.Types;
 using ScadaLink.Commons.Types.Enums;
 
 namespace ScadaLink.AuditLog.Site;
@@ -390,6 +390,184 @@ public class SqliteAuditWriter : IAuditWriter, ISiteAuditQueue, IAsyncDisposable
         }
     }
 
+    /// <summary>
+    /// M6 reconciliation-pull read: returns up to <paramref name="batchSize"/> rows
+    /// whose <c>OccurredAtUtc &gt;= sinceUtc</c> and whose <see cref="AuditForwardState"/>
+    /// is still <see cref="AuditForwardState.Pending"/> or
+    /// <see cref="AuditForwardState.Forwarded"/>. Forwarded rows are included so the
+    /// brief race window between a site-Forwarded ack and central ingest cannot
+    /// silently drop rows; central dedups on <see cref="AuditEvent.EventId"/>.
+    /// Ordered oldest <see cref="AuditEvent.OccurredAtUtc"/> first, EventId tiebreaker.
+    /// </summary>
+    public Task<IReadOnlyList<AuditEvent>> ReadPendingSinceAsync(
+        DateTime sinceUtc, int batchSize, CancellationToken ct = default)
+    {
+        if (batchSize <= 0)
+        {
+            throw new ArgumentOutOfRangeException(nameof(batchSize), "batchSize must be > 0.");
+        }
+
+        // Mirror ReadPendingAsync: the write lock guards the single connection.
+        lock (_writeLock)
+        {
+            ObjectDisposedException.ThrowIf(_disposed, this);
+
+            using var cmd = _connection.CreateCommand();
+            cmd.CommandText = """
+                SELECT EventId, OccurredAtUtc, Channel, Kind, CorrelationId,
+                       SourceSiteId, SourceInstanceId, SourceScript, Actor, Target,
+                       Status, HttpStatus, DurationMs, ErrorMessage, ErrorDetail,
+                       RequestSummary, ResponseSummary, PayloadTruncated, Extra, ForwardState
+                FROM   AuditLog
+                WHERE  ForwardState IN ($pending, $forwarded)
+                  AND  OccurredAtUtc >= $since
+                ORDER  BY OccurredAtUtc ASC, EventId ASC
+                LIMIT  $limit;
+                """;
+            cmd.Parameters.AddWithValue("$pending", AuditForwardState.Pending.ToString());
+            cmd.Parameters.AddWithValue("$forwarded", AuditForwardState.Forwarded.ToString());
+            // Normalise to UTC ISO-8601 round-trip format to match how OccurredAtUtc
+            // is stored on insert ("o" format) — string comparison is monotonic for
+            // that encoding so we can index-scan against it.
+            cmd.Parameters.AddWithValue("$since", EnsureUtc(sinceUtc).ToString(
+                "o", System.Globalization.CultureInfo.InvariantCulture));
+            cmd.Parameters.AddWithValue("$limit", batchSize);
+
+            var rows = new List<AuditEvent>(Math.Min(batchSize, 256));
+            using var reader = cmd.ExecuteReader();
+            while (reader.Read())
+            {
+                rows.Add(MapRow(reader));
+            }
+
+            return Task.FromResult<IReadOnlyList<AuditEvent>>(rows);
+        }
+    }
+
+    /// <summary>
+    /// M6 reconciliation-pull commit: flips the supplied EventIds to
+    /// <see cref="AuditForwardState.Reconciled"/>, but ONLY for rows currently in
+    /// <see cref="AuditForwardState.Pending"/> or <see cref="AuditForwardState.Forwarded"/>.
+    /// Rows already in <see cref="AuditForwardState.Reconciled"/> are left untouched
+    /// (idempotent re-call). Non-existent ids are silent no-ops.
+    /// </summary>
+    public Task MarkReconciledAsync(IReadOnlyList<Guid> eventIds, CancellationToken ct = default)
+    {
+        ArgumentNullException.ThrowIfNull(eventIds);
+        if (eventIds.Count == 0)
+        {
+            return Task.CompletedTask;
+        }
+
+        lock (_writeLock)
+        {
+            ObjectDisposedException.ThrowIf(_disposed, this);
+
+            using var cmd = _connection.CreateCommand();
+            var sb = new System.Text.StringBuilder();
+            sb.Append("UPDATE AuditLog SET ForwardState = $reconciled ")
+              .Append("WHERE ForwardState IN ($pending, $forwarded) AND EventId IN (");
+            for (int i = 0; i < eventIds.Count; i++)
+            {
+                if (i > 0) sb.Append(',');
+                var p = $"$id{i}";
+                sb.Append(p);
+                cmd.Parameters.AddWithValue(p, eventIds[i].ToString());
+            }
+            sb.Append(");");
+            cmd.CommandText = sb.ToString();
+            cmd.Parameters.AddWithValue("$reconciled", AuditForwardState.Reconciled.ToString());
+            cmd.Parameters.AddWithValue("$pending", AuditForwardState.Pending.ToString());
+            cmd.Parameters.AddWithValue("$forwarded", AuditForwardState.Forwarded.ToString());
+
+            cmd.ExecuteNonQuery();
+            return Task.CompletedTask;
+        }
+    }
+
+    /// <summary>
+    /// M6 Bundle E (T6) health-metric surface: returns a point-in-time snapshot
+    /// of the site queue's pending count, the oldest pending row's
+    /// <see cref="AuditEvent.OccurredAtUtc"/>, and the on-disk file size. Called
+    /// by the site-side <c>SiteAuditBacklogReporter</c> hosted service on its
+    /// 30 s tick to refresh the <c>SiteHealthReport.SiteAuditBacklog</c> field.
+    /// </summary>
+    /// <remarks>
+    /// The pending-count + oldest-row queries run inside the same write lock as
+    /// the hot-path INSERT batch so the snapshot is consistent against the
+    /// connection's view (no torn read of an in-flight transaction). The on-disk
+    /// size lookup happens OUTSIDE the lock — it's a stat() call on the file
+    /// path and doesn't touch the connection. In-memory and missing files
+    /// return 0 bytes (the snapshot is for ops dashboards, not a correctness
+    /// invariant).
+    /// </remarks>
+    public Task<SiteAuditBacklogSnapshot> GetBacklogStatsAsync(CancellationToken ct = default)
+    {
+        int pendingCount;
+        DateTime? oldestPending;
+
+        lock (_writeLock)
+        {
+            ObjectDisposedException.ThrowIf(_disposed, this);
+
+            // Single round-trip — COUNT(*) + MIN(OccurredAtUtc) over the same
+            // index range avoids a second scan. The IX_SiteAuditLog_ForwardState_Occurred
+            // index makes both aggregates cheap (count is a covering scan, min
+            // is the first key).
+            using var cmd = _connection.CreateCommand();
+            cmd.CommandText = """
+                SELECT COUNT(*), MIN(OccurredAtUtc)
+                FROM   AuditLog
+                WHERE  ForwardState = $pending;
+                """;
+            cmd.Parameters.AddWithValue("$pending", AuditForwardState.Pending.ToString());
+
+            using var reader = cmd.ExecuteReader();
+            reader.Read();
+            pendingCount = reader.GetInt32(0);
+            oldestPending = reader.IsDBNull(1)
+                ? null
+                : DateTime.Parse(reader.GetString(1),
+                    System.Globalization.CultureInfo.InvariantCulture,
+                    System.Globalization.DateTimeStyles.RoundtripKind);
+        }
+
+        // File-size lookup outside the lock — the DatabasePath option is the
+        // canonical source. The connection-string-override branch (used by
+        // some tests) keeps the same DatabasePath value, so this works
+        // uniformly. In-memory / mode=memory paths return 0 because the file
+        // doesn't exist on disk.
+        long onDiskBytes = 0;
+        try
+        {
+            if (!string.IsNullOrEmpty(_options.DatabasePath) &&
+                !_options.DatabasePath.StartsWith(":memory:", StringComparison.Ordinal) &&
+                !_options.DatabasePath.Contains("mode=memory", StringComparison.OrdinalIgnoreCase) &&
+                File.Exists(_options.DatabasePath))
+            {
+                onDiskBytes = new FileInfo(_options.DatabasePath).Length;
+            }
+        }
+        catch (Exception ex)
+        {
+            // File system probe is a best-effort health-metric — never abort
+            // a backlog snapshot because stat() failed. Log and report 0.
+            _logger.LogDebug(ex,
+                "SqliteAuditWriter could not stat DB path {Path} for backlog snapshot.",
+                _options.DatabasePath);
+        }
+
+        return Task.FromResult(new SiteAuditBacklogSnapshot(
+            PendingCount: pendingCount,
+            OldestPendingUtc: oldestPending,
+            OnDiskBytes: onDiskBytes));
+    }
+
+    private static DateTime EnsureUtc(DateTime value) =>
+        value.Kind == DateTimeKind.Utc
+            ? value
+            : DateTime.SpecifyKind(value.ToUniversalTime(), DateTimeKind.Utc);
+
     private static AuditEvent MapRow(SqliteDataReader reader)
     {
         return new AuditEvent

src/ScadaLink.AuditLog/Site/Telemetry/ISiteAuditQueue.cs

@@ -1,34 +0,0 @@
-using ScadaLink.Commons.Entities.Audit;
-
-namespace ScadaLink.AuditLog.Site.Telemetry;
-
-/// <summary>
-/// Site-local audit-log queue surface consumed by <see cref="SiteAuditTelemetryActor"/>.
-/// Extracted from <see cref="SqliteAuditWriter"/> so the telemetry actor can be
-/// unit-tested against a stub without touching SQLite. <see cref="SqliteAuditWriter"/>
-/// implements this interface; production wiring injects the same instance.
-/// </summary>
-/// <remarks>
-/// Only the two methods the drain loop needs are exposed — the hot-path
-/// <c>WriteAsync</c> stays on <see cref="Commons.Interfaces.Services.IAuditWriter"/>
-/// (script-thread surface), separated by concern from the
-/// telemetry-actor surface so each side can be mocked independently.
-/// </remarks>
-public interface ISiteAuditQueue
-{
-    /// <summary>
-    /// Returns up to <paramref name="limit"/> rows currently in
-    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/>,
-    /// oldest first. Idempotent — repeated calls before
-    /// <see cref="MarkForwardedAsync"/> will yield the same rows again.
-    /// </summary>
-    Task<IReadOnlyList<AuditEvent>> ReadPendingAsync(int limit, CancellationToken ct = default);
-
-    /// <summary>
-    /// Flips the supplied EventIds from
-    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/> to
-    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Forwarded"/>.
-    /// Non-existent or already-forwarded ids are silent no-ops.
-    /// </summary>
-    Task MarkForwardedAsync(IReadOnlyList<Guid> eventIds, CancellationToken ct = default);
-}

src/ScadaLink.AuditLog/Site/Telemetry/SiteAuditTelemetryActor.cs

@@ -3,6 +3,7 @@ using Microsoft.Extensions.Logging;
 using Microsoft.Extensions.Options;
 using ScadaLink.AuditLog.Telemetry;
 using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Services;
 using ScadaLink.Communication.Grpc;
 
 namespace ScadaLink.AuditLog.Site.Telemetry;

src/ScadaLink.Commons/Interfaces/IPartitionMaintenance.cs

@@ -0,0 +1,48 @@
+namespace ScadaLink.Commons.Interfaces;
+
+/// <summary>
+/// Abstraction over the central AuditLog partition-function roll-forward
+/// operation. M6-T5 introduces a daily-cadence hosted service
+/// (<c>AuditLogPartitionMaintenanceService</c>) that calls
+/// <see cref="EnsureLookaheadAsync"/> to make sure
+/// <c>pf_AuditLog_Month</c> always has at least <c>LookaheadMonths</c> of
+/// future boundaries available — otherwise inserts past the highest
+/// boundary land in a single ever-growing tail partition that
+/// <c>SwitchOutPartitionAsync</c> cannot purge cleanly.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The interface lives in <c>ScadaLink.Commons</c> so the central hosted
+/// service in <c>ScadaLink.AuditLog</c> can depend on it without taking a
+/// reference on <c>ScadaLink.ConfigurationDatabase</c>; the EF-based
+/// implementation ships in
+/// <c>ScadaLink.ConfigurationDatabase.Maintenance.AuditLogPartitionMaintenance</c>
+/// and is registered by <c>AddConfigurationDatabase</c>.
+/// </para>
+/// <para>
+/// Both methods read <c>sys.partition_range_values</c> / mutate
+/// <c>pf_AuditLog_Month</c> via raw SQL — there is no EF model for a
+/// partition function. The interface deliberately exposes only the two
+/// operations the hosted service needs; it is not a general partition-DDL
+/// surface.
+/// </para>
+/// </remarks>
+public interface IPartitionMaintenance
+{
+    /// <summary>
+    /// Splits new monthly boundaries on <c>pf_AuditLog_Month</c> so the
+    /// function covers at least <paramref name="lookaheadMonths"/> future
+    /// months relative to <see cref="DateTime.UtcNow"/>. Idempotent — a
+    /// boundary that already exists is skipped rather than re-issued.
+    /// Returns the boundaries actually added, in chronological order.
+    /// </summary>
+    Task<IReadOnlyList<DateTime>> EnsureLookaheadAsync(int lookaheadMonths, CancellationToken ct = default);
+
+    /// <summary>
+    /// Reads the current maximum boundary value from
+    /// <c>sys.partition_range_values</c> for <c>pf_AuditLog_Month</c>.
+    /// Returns <c>null</c> when the partition function does not exist or
+    /// has no boundaries.
+    /// </summary>
+    Task<DateTime?> GetMaxBoundaryAsync(CancellationToken ct = default);
+}

src/ScadaLink.Commons/Interfaces/Repositories/IAuditLogRepository.cs

@@ -45,12 +45,46 @@ public interface IAuditLogRepository
 
     /// <summary>
     /// Switches out (purges) the monthly partition whose lower bound is
-    /// <paramref name="monthBoundary"/>. The honest M1 implementation throws
-    /// <see cref="NotSupportedException"/>: the <c>UX_AuditLog_EventId</c> unique
-    /// index is non-partition-aligned (lives on <c>[PRIMARY]</c>, not on
-    /// <c>ps_AuditLog_Month</c>), so SQL Server rejects
-    /// <c>ALTER TABLE … SWITCH PARTITION</c> until the drop-and-rebuild dance
-    /// shipped by the M6 purge actor is in place.
+    /// <paramref name="monthBoundary"/> and returns the approximate number
+    /// of rows discarded — sampled inside the transaction BEFORE the switch
+    /// so the row count reflects what the switch removed, not a post-purge
+    /// scan of a table that no longer exists.
     /// </summary>
-    Task SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default);
+    /// <remarks>
+    /// <para>
+    /// <b>Drop-and-rebuild dance.</b> <c>UX_AuditLog_EventId</c> is intentionally
+    /// non-partition-aligned (it lives on <c>[PRIMARY]</c> so single-column
+    /// EventId uniqueness — required by <see cref="InsertIfNotExistsAsync"/> —
+    /// can be enforced cheaply). SQL Server rejects
+    /// <c>ALTER TABLE … SWITCH PARTITION</c> while a non-aligned unique index
+    /// is present, so the M6 implementation drops the index, creates a staging
+    /// table with byte-identical schema, switches the partition's data into
+    /// staging, drops staging (discarding the rows), and rebuilds the unique
+    /// index. The CATCH branch guarantees the index is rebuilt even on partial
+    /// failure so the table never returns to live traffic without its
+    /// idempotency-supporting index.
+    /// </para>
+    /// <para>
+    /// <b>Outage window.</b> The dance briefly removes the unique index, so
+    /// concurrent <see cref="InsertIfNotExistsAsync"/> calls during the switch
+    /// could in principle race past the IF NOT EXISTS check without the index
+    /// catching the duplicate. This is acceptable for the daily purge cadence
+    /// — the inserts that the IF NOT EXISTS check guards are themselves rare
+    /// enough that a sub-second collision window is operationally negligible,
+    /// and the composite PK still rejects same-(EventId, OccurredAtUtc) rows.
+    /// </para>
+    /// </remarks>
+    Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default);
+
+    /// <summary>
+    /// Returns the set of <c>pf_AuditLog_Month</c> partition lower-bound
+    /// boundaries whose partitions contain only rows with
+    /// <see cref="AuditEvent.OccurredAtUtc"/> strictly older than
+    /// <paramref name="threshold"/>. Boundaries whose partition is empty are
+    /// excluded (a no-op switch is wasted work). Used by the M6 purge actor
+    /// to enumerate retention-eligible months on every tick.
+    /// </summary>
+    Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+        DateTime threshold,
+        CancellationToken ct = default);
 }

src/ScadaLink.Commons/Interfaces/Services/ISiteAuditQueue.cs

@@ -0,0 +1,87 @@
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Types;
+
+namespace ScadaLink.Commons.Interfaces.Services;
+
+/// <summary>
+/// Site-local audit-log queue surface consumed by the site
+/// <c>SiteAuditTelemetryActor</c> drain loop and the M6
+/// <c>SiteStreamGrpcServer.PullAuditEvents</c> reconciliation handler.
+/// Extracted from <c>SqliteAuditWriter</c> so both consumers can be
+/// unit-tested against a stub without touching SQLite; the
+/// <c>SqliteAuditWriter</c> production type implements this interface
+/// and DI wires the same singleton instance to every consumer.
+/// </summary>
+/// <remarks>
+/// Lives in Commons (rather than alongside <c>SqliteAuditWriter</c> in
+/// <c>ScadaLink.AuditLog</c>) because <c>ScadaLink.Communication</c> — which
+/// hosts the M6 gRPC pull handler — must depend on this interface and
+/// <c>ScadaLink.AuditLog</c> already depends on <c>ScadaLink.Communication</c>.
+/// Pulling the interface up to Commons breaks the would-be cycle while
+/// keeping the implementation in the AuditLog component.
+///
+/// Only the methods the drain and pull paths need are exposed — the
+/// hot-path <c>WriteAsync</c> stays on <see cref="IAuditWriter"/>
+/// (script-thread surface), separated by concern so each side can be
+/// mocked independently.
+/// </remarks>
+public interface ISiteAuditQueue
+{
+    /// <summary>
+    /// Returns up to <paramref name="limit"/> rows currently in
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/>,
+    /// oldest first. Idempotent — repeated calls before
+    /// <see cref="MarkForwardedAsync"/> will yield the same rows again.
+    /// </summary>
+    Task<IReadOnlyList<AuditEvent>> ReadPendingAsync(int limit, CancellationToken ct = default);
+
+    /// <summary>
+    /// Flips the supplied EventIds from
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/> to
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Forwarded"/>.
+    /// Non-existent or already-forwarded ids are silent no-ops.
+    /// </summary>
+    Task MarkForwardedAsync(IReadOnlyList<Guid> eventIds, CancellationToken ct = default);
+
+    /// <summary>
+    /// M6 reconciliation-pull read surface: returns up to <paramref name="batchSize"/>
+    /// rows whose <see cref="AuditEvent.OccurredAtUtc"/> &gt;= <paramref name="sinceUtc"/>
+    /// and whose <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState"/> is still
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/> or
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Forwarded"/>.
+    /// </summary>
+    /// <remarks>
+    /// Rows in the brief race window between site-Forwarded and central-ingest are
+    /// intentionally included: the central reconciliation puller dedups on
+    /// <see cref="AuditEvent.EventId"/>, so re-shipping is safe and avoids losing rows
+    /// whose telemetry ack was acted on locally but never landed centrally. Ordering
+    /// is oldest <see cref="AuditEvent.OccurredAtUtc"/> first with
+    /// <see cref="AuditEvent.EventId"/> as the deterministic tiebreaker.
+    /// </remarks>
+    Task<IReadOnlyList<AuditEvent>> ReadPendingSinceAsync(
+        DateTime sinceUtc, int batchSize, CancellationToken ct = default);
+
+    /// <summary>
+    /// M6 reconciliation-pull commit surface: flips the supplied EventIds to
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Reconciled"/>,
+    /// but ONLY for rows currently in
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/> or
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Forwarded"/>.
+    /// Rows already in <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Reconciled"/>
+    /// are left untouched (idempotent re-call). Non-existent ids are silent no-ops.
+    /// </summary>
+    Task MarkReconciledAsync(IReadOnlyList<Guid> eventIds, CancellationToken ct = default);
+
+    /// <summary>
+    /// M6 Bundle E (T6) health-metric surface: returns a point-in-time snapshot
+    /// of the site queue's pending count + oldest pending timestamp + on-disk
+    /// SQLite file size. Surfaced on
+    /// <see cref="ScadaLink.Commons.Messages.Health.SiteHealthReport"/> as
+    /// <c>SiteAuditBacklog</c> by the periodic <c>SiteAuditBacklogReporter</c>
+    /// hosted service so a stuck site→central drain is visible on the central
+    /// health dashboard. Safe to call concurrently with hot-path writes —
+    /// implementations are expected to take the same connection lock used by
+    /// the hot-path INSERT batch and the drain queries.
+    /// </summary>
+    Task<SiteAuditBacklogSnapshot> GetBacklogStatsAsync(CancellationToken ct = default);
+}

src/ScadaLink.Commons/Messages/Health/SiteHealthReport.cs

@@ -1,3 +1,4 @@
+using ScadaLink.Commons.Types;
 using ScadaLink.Commons.Types.Enums;
 
 namespace ScadaLink.Commons.Messages.Health;
@@ -32,7 +33,14 @@ public record SiteHealthReport(
     // marker). Surfaces a misconfigured / catastrophic regex on
     // /monitoring/health. Defaults to 0 for back-compat with existing
     // producers and tests that don't construct the field.
-    int AuditRedactionFailure = 0);
+    int AuditRedactionFailure = 0,
+    // Audit Log (#23) M6 Bundle E (T6): point-in-time snapshot of the
+    // site-local SQLite audit-log queue (pending count, oldest pending row,
+    // on-disk bytes). Populated by the site-side SiteAuditBacklogReporter
+    // hosted service every 30 s. Defaults to null so existing producers /
+    // tests that don't refresh the snapshot stay valid; the central health
+    // surface treats null as "no data yet" rather than a zeroed queue.
+    SiteAuditBacklogSnapshot? SiteAuditBacklog = null);
 
 /// <summary>
 /// Broadcast wrapper used between central nodes to keep per-node

src/ScadaLink.Commons/Types/SiteAuditBacklogSnapshot.cs

@@ -0,0 +1,32 @@
+namespace ScadaLink.Commons.Types;
+
+/// <summary>
+/// Audit Log (#23) M6 Bundle E (T6) — point-in-time snapshot of the site-local
+/// SQLite audit-log queue health, surfaced on
+/// <see cref="ScadaLink.Commons.Messages.Health.SiteHealthReport"/> as
+/// <c>SiteAuditBacklog</c> and refreshed periodically by the
+/// <c>SiteAuditBacklogReporter</c> hosted service.
+/// </summary>
+/// <param name="PendingCount">
+/// Number of rows currently in
+/// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Pending"/> — i.e.
+/// not yet acknowledged by central via either the push-telemetry or
+/// reconciliation-pull paths. A persistently non-zero value with rising
+/// <see cref="OldestPendingUtc"/> indicates the site→central drain isn't
+/// keeping up.
+/// </param>
+/// <param name="OldestPendingUtc">
+/// <see cref="ScadaLink.Commons.Entities.Audit.AuditEvent.OccurredAtUtc"/> of
+/// the oldest Pending row, or <c>null</c> if the queue is empty. Used by ops
+/// to compute backlog age without a separate query.
+/// </param>
+/// <param name="OnDiskBytes">
+/// Size of the SQLite file on disk in bytes, or <c>0</c> if the writer is
+/// running against an in-memory database. Mirrors the 7-day retention
+/// invariant (alog.md §10) — a steady file-size growth past the retention
+/// window points at a stuck purge or a stuck forwarder.
+/// </param>
+public sealed record SiteAuditBacklogSnapshot(
+    int PendingCount,
+    DateTime? OldestPendingUtc,
+    long OnDiskBytes);

src/ScadaLink.Communication/Grpc/SiteStreamGrpcServer.cs

@@ -5,6 +5,7 @@ using Grpc.Core;
 using Microsoft.Extensions.Logging;
 using Microsoft.Extensions.Options;
 using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Services;
 using ScadaLink.Commons.Messages.Audit;
 using ScadaLink.Commons.Types;
 using ScadaLink.Commons.Types.Enums;
@@ -36,6 +37,13 @@ public class SiteStreamGrpcServer : SiteStreamService.SiteStreamServiceBase
     // calls are sub-100 ms in steady state; a generous timeout absorbs a slow
     // MSSQL connection without surfacing as a gRPC failure on a healthy site.
     private static readonly TimeSpan AuditIngestAskTimeout = TimeSpan.FromSeconds(30);
+    // Audit Log (#23 M6): site-local queue handed in by AkkaHostedService on
+    // site roles so the central reconciliation puller's PullAuditEvents RPC
+    // can read Pending/Forwarded rows. Null when not wired (e.g. central-only
+    // host or test composing the server in isolation) — the handler treats
+    // the missing queue as "nothing to ship" and returns an empty response so
+    // central retries on its next reconciliation cycle.
+    private ISiteAuditQueue? _siteAuditQueue;
 
     /// <summary>
     /// Test-only constructor — kept <c>internal</c> so the DI container sees a
@@ -102,6 +110,20 @@ public class SiteStreamGrpcServer : SiteStreamService.SiteStreamServiceBase
         _auditIngestActor = proxy;
     }
 
+    /// <summary>
+    /// Hands the site-local <see cref="ISiteAuditQueue"/> (the same
+    /// <c>SqliteAuditWriter</c> singleton that backs <see cref="IAuditWriter"/>
+    /// on the script thread) to the gRPC server so the M6
+    /// <see cref="PullAuditEvents"/> RPC can serve central's reconciliation
+    /// pulls. Mirrors <see cref="SetAuditIngestActor"/>: wired post-construction
+    /// because the queue and the gRPC server are both DI singletons brought up
+    /// in independent orders on site startup.
+    /// </summary>
+    public void SetSiteAuditQueue(ISiteAuditQueue queue)
+    {
+        _siteAuditQueue = queue;
+    }
+
     /// <summary>
     /// Number of currently active streaming subscriptions. Exposed for diagnostics.
     /// </summary>
@@ -361,6 +383,144 @@ public class SiteStreamGrpcServer : SiteStreamService.SiteStreamServiceBase
         return ack;
     }
 
+    /// <summary>
+    /// Audit Log (#23) M6 reconciliation pull RPC. Central asks the site for any
+    /// AuditLog rows whose <c>OccurredAtUtc &gt;= since_utc</c> and whose
+    /// <c>ForwardState</c> is still <c>Pending</c> or <c>Forwarded</c> (i.e. not
+    /// yet confirmed reconciled), bounded by <c>batch_size</c>. The site responds
+    /// with the rows AND flips them to
+    /// <see cref="ScadaLink.Commons.Types.Enums.AuditForwardState.Reconciled"/>
+    /// AFTER serializing the response. The flip is best-effort — if it fails
+    /// (e.g. SQLite disposed mid-call), rows stay Pending/Forwarded and central
+    /// pulls them again on the next reconciliation cycle. Idempotent.
+    /// </summary>
+    /// <remarks>
+    /// When <see cref="_siteAuditQueue"/> is not wired (central-only host or a
+    /// composition-root test exercising the server in isolation) the RPC returns
+    /// an empty response — central treats that as "nothing to ship" and retries
+    /// on its next cycle, which is the same self-healing semantics as the
+    /// SetAuditIngestActor wiring race window.
+    /// </remarks>
+    public override async Task<PullAuditEventsResponse> PullAuditEvents(
+        PullAuditEventsRequest request,
+        ServerCallContext context)
+    {
+        var queue = _siteAuditQueue;
+        if (queue is null)
+        {
+            _logger.LogWarning(
+                "PullAuditEvents invoked before SetSiteAuditQueue was called; returning empty response.");
+            return new PullAuditEventsResponse();
+        }
+
+        if (request.BatchSize <= 0)
+        {
+            // Mirrors the SubscribeInstance guard: reject malformed requests
+            // cleanly with InvalidArgument so the caller doesn't see a generic
+            // RpcException from the underlying SQLite parameter validation.
+            throw new RpcException(new GrpcStatus(
+                StatusCode.InvalidArgument, "batch_size must be > 0"));
+        }
+
+        // sinceUtc defaults to DateTime.MinValue when the wrapper is absent —
+        // i.e. "pull from the beginning of recorded history", which is the
+        // intended behaviour for the very first reconciliation cycle.
+        var since = request.SinceUtc?.ToDateTime().ToUniversalTime() ?? DateTime.MinValue;
+
+        IReadOnlyList<AuditEvent> events;
+        try
+        {
+            events = await queue.ReadPendingSinceAsync(
+                since, request.BatchSize, context.CancellationToken);
+        }
+        catch (Exception ex)
+        {
+            _logger.LogError(ex,
+                "ReadPendingSinceAsync failed for since={Since} batch={Batch}; returning empty response.",
+                since, request.BatchSize);
+            return new PullAuditEventsResponse();
+        }
+
+        var response = new PullAuditEventsResponse
+        {
+            // batch_size saturated → tell central to issue a follow-up pull
+            // with an advanced cursor. The site doesn't compute the cursor —
+            // central walks it forward from the last returned OccurredAtUtc.
+            MoreAvailable = events.Count >= request.BatchSize,
+        };
+        foreach (var evt in events)
+        {
+            response.Events.Add(AuditEventToDto(evt));
+        }
+
+        // Flip to Reconciled AFTER projecting the response so a fault below the
+        // try/catch (mid-response, mid-flip) leaves the rows in Pending/Forwarded
+        // and central pulls them again next cycle. The flip itself is
+        // best-effort — its failure is a warning, not a fault, because central
+        // will dedup on EventId on the next pull.
+        var ids = new List<Guid>(events.Count);
+        foreach (var evt in events)
+        {
+            ids.Add(evt.EventId);
+        }
+
+        if (ids.Count > 0)
+        {
+            try
+            {
+                await queue.MarkReconciledAsync(ids, context.CancellationToken);
+            }
+            catch (Exception ex)
+            {
+                _logger.LogWarning(ex,
+                    "MarkReconciledAsync failed after PullAuditEvents response of {Count} rows; rows stay Pending for retry.",
+                    ids.Count);
+            }
+        }
+
+        return response;
+    }
+
+    /// <summary>
+    /// Inlined audit-event entity→DTO translation. Keep in sync with
+    /// <c>AuditEventMapper.ToDto</c> in <c>ScadaLink.AuditLog.Telemetry</c> —
+    /// the project-reference cycle (AuditLog → Communication) prevents calling
+    /// the AuditLog mapper directly. The shape mirrors the FromDto pair above.
+    /// </summary>
+    private static AuditEventDto AuditEventToDto(AuditEvent evt)
+    {
+        var dto = new AuditEventDto
+        {
+            EventId = evt.EventId.ToString(),
+            OccurredAtUtc = Google.Protobuf.WellKnownTypes.Timestamp.FromDateTime(EnsureUtc(evt.OccurredAtUtc)),
+            Channel = evt.Channel.ToString(),
+            Kind = evt.Kind.ToString(),
+            CorrelationId = evt.CorrelationId?.ToString() ?? string.Empty,
+            SourceSiteId = evt.SourceSiteId ?? string.Empty,
+            SourceInstanceId = evt.SourceInstanceId ?? string.Empty,
+            SourceScript = evt.SourceScript ?? string.Empty,
+            Actor = evt.Actor ?? string.Empty,
+            Target = evt.Target ?? string.Empty,
+            Status = evt.Status.ToString(),
+            ErrorMessage = evt.ErrorMessage ?? string.Empty,
+            ErrorDetail = evt.ErrorDetail ?? string.Empty,
+            RequestSummary = evt.RequestSummary ?? string.Empty,
+            ResponseSummary = evt.ResponseSummary ?? string.Empty,
+            PayloadTruncated = evt.PayloadTruncated,
+            Extra = evt.Extra ?? string.Empty,
+        };
+
+        if (evt.HttpStatus.HasValue) dto.HttpStatus = evt.HttpStatus.Value;
+        if (evt.DurationMs.HasValue) dto.DurationMs = evt.DurationMs.Value;
+
+        return dto;
+    }
+
+    private static DateTime EnsureUtc(DateTime value) =>
+        value.Kind == DateTimeKind.Utc
+            ? value
+            : DateTime.SpecifyKind(value.ToUniversalTime(), DateTimeKind.Utc);
+
     private static string? NullIfEmpty(string? value) =>
         string.IsNullOrEmpty(value) ? null : value;
 

src/ScadaLink.Communication/Protos/sitestream.proto

@@ -9,6 +9,7 @@ service SiteStreamService {
   rpc SubscribeInstance(InstanceStreamRequest) returns (stream SiteStreamEvent);
   rpc IngestAuditEvents(AuditEventBatch) returns (IngestAck);
   rpc IngestCachedTelemetry(CachedTelemetryBatch) returns (IngestAck);
+  rpc PullAuditEvents(PullAuditEventsRequest) returns (PullAuditEventsResponse);
 }
 
 message InstanceStreamRequest {
@@ -119,3 +120,19 @@ message CachedTelemetryPacket {
 }
 
 message CachedTelemetryBatch { repeated CachedTelemetryPacket packets = 1; }
+
+// Audit Log (#23) M6 reconciliation pull: central→site request for any
+// site-local AuditLog rows with OccurredAtUtc >= since_utc that have not yet
+// been ingested centrally (ForwardState in {Pending, Forwarded}). The site
+// flips returned rows to Reconciled after the response is on the wire.
+// more_available signals batch_size was saturated so the caller knows to
+// issue a follow-up pull with an advanced since_utc cursor.
+message PullAuditEventsRequest {
+  google.protobuf.Timestamp since_utc = 1;
+  int32 batch_size = 2;
+}
+
+message PullAuditEventsResponse {
+  repeated AuditEventDto events = 1;
+  bool more_available = 2;
+}

src/ScadaLink.Communication/SiteStreamGrpc/Sitestream.cs

@@ -68,21 +68,27 @@ namespace ScadaLink.Communication.Grpc {
             "bnREdG8SNwoLb3BlcmF0aW9uYWwYAiABKAsyIi5zaXRlc3RyZWFtLlNpdGVD",
             "YWxsT3BlcmF0aW9uYWxEdG8iSgoUQ2FjaGVkVGVsZW1ldHJ5QmF0Y2gSMgoH",
             "cGFja2V0cxgBIAMoCzIhLnNpdGVzdHJlYW0uQ2FjaGVkVGVsZW1ldHJ5UGFj",
-            "a2V0KlwKB1F1YWxpdHkSFwoTUVVBTElUWV9VTlNQRUNJRklFRBAAEhAKDFFV",
-            "QUxJVFlfR09PRBABEhUKEVFVQUxJVFlfVU5DRVJUQUlOEAISDwoLUVVBTElU",
-            "WV9CQUQQAypdCg5BbGFybVN0YXRlRW51bRIbChdBTEFSTV9TVEFURV9VTlNQ",
-            "RUNJRklFRBAAEhYKEkFMQVJNX1NUQVRFX05PUk1BTBABEhYKEkFMQVJNX1NU",
-            "QVRFX0FDVElWRRACKoUBCg5BbGFybUxldmVsRW51bRIUChBBTEFSTV9MRVZF",
-            "TF9OT05FEAASEwoPQUxBUk1fTEVWRUxfTE9XEAESFwoTQUxBUk1fTEVWRUxf",
-            "TE9XX0xPVxACEhQKEEFMQVJNX0xFVkVMX0hJR0gQAxIZChVBTEFSTV9MRVZF",
-            "TF9ISUdIX0hJR0gQBDKFAgoRU2l0ZVN0cmVhbVNlcnZpY2USVQoRU3Vic2Ny",
-            "aWJlSW5zdGFuY2USIS5zaXRlc3RyZWFtLkluc3RhbmNlU3RyZWFtUmVxdWVz",
-            "dBobLnNpdGVzdHJlYW0uU2l0ZVN0cmVhbUV2ZW50MAESRwoRSW5nZXN0QXVk",
-            "aXRFdmVudHMSGy5zaXRlc3RyZWFtLkF1ZGl0RXZlbnRCYXRjaBoVLnNpdGVz",
-            "dHJlYW0uSW5nZXN0QWNrElAKFUluZ2VzdENhY2hlZFRlbGVtZXRyeRIgLnNp",
-            "dGVzdHJlYW0uQ2FjaGVkVGVsZW1ldHJ5QmF0Y2gaFS5zaXRlc3RyZWFtLklu",
-            "Z2VzdEFja0IfqgIcU2NhZGFMaW5rLkNvbW11bmljYXRpb24uR3JwY2IGcHJv",
-            "dG8z"));
+            "a2V0IlsKFlB1bGxBdWRpdEV2ZW50c1JlcXVlc3QSLQoJc2luY2VfdXRjGAEg",
+            "ASgLMhouZ29vZ2xlLnByb3RvYnVmLlRpbWVzdGFtcBISCgpiYXRjaF9zaXpl",
+            "GAIgASgFIlwKF1B1bGxBdWRpdEV2ZW50c1Jlc3BvbnNlEikKBmV2ZW50cxgB",
+            "IAMoCzIZLnNpdGVzdHJlYW0uQXVkaXRFdmVudER0bxIWCg5tb3JlX2F2YWls",
+            "YWJsZRgCIAEoCCpcCgdRdWFsaXR5EhcKE1FVQUxJVFlfVU5TUEVDSUZJRUQQ",
+            "ABIQCgxRVUFMSVRZX0dPT0QQARIVChFRVUFMSVRZX1VOQ0VSVEFJThACEg8K",
+            "C1FVQUxJVFlfQkFEEAMqXQoOQWxhcm1TdGF0ZUVudW0SGwoXQUxBUk1fU1RB",
+            "VEVfVU5TUEVDSUZJRUQQABIWChJBTEFSTV9TVEFURV9OT1JNQUwQARIWChJB",
+            "TEFSTV9TVEFURV9BQ1RJVkUQAiqFAQoOQWxhcm1MZXZlbEVudW0SFAoQQUxB",
+            "Uk1fTEVWRUxfTk9ORRAAEhMKD0FMQVJNX0xFVkVMX0xPVxABEhcKE0FMQVJN",
+            "X0xFVkVMX0xPV19MT1cQAhIUChBBTEFSTV9MRVZFTF9ISUdIEAMSGQoVQUxB",
+            "Uk1fTEVWRUxfSElHSF9ISUdIEAQy4QIKEVNpdGVTdHJlYW1TZXJ2aWNlElUK",
+            "EVN1YnNjcmliZUluc3RhbmNlEiEuc2l0ZXN0cmVhbS5JbnN0YW5jZVN0cmVh",
+            "bVJlcXVlc3QaGy5zaXRlc3RyZWFtLlNpdGVTdHJlYW1FdmVudDABEkcKEUlu",
+            "Z2VzdEF1ZGl0RXZlbnRzEhsuc2l0ZXN0cmVhbS5BdWRpdEV2ZW50QmF0Y2ga",
+            "FS5zaXRlc3RyZWFtLkluZ2VzdEFjaxJQChVJbmdlc3RDYWNoZWRUZWxlbWV0",
+            "cnkSIC5zaXRlc3RyZWFtLkNhY2hlZFRlbGVtZXRyeUJhdGNoGhUuc2l0ZXN0",
+            "cmVhbS5Jbmdlc3RBY2sSWgoPUHVsbEF1ZGl0RXZlbnRzEiIuc2l0ZXN0cmVh",
+            "bS5QdWxsQXVkaXRFdmVudHNSZXF1ZXN0GiMuc2l0ZXN0cmVhbS5QdWxsQXVk",
+            "aXRFdmVudHNSZXNwb25zZUIfqgIcU2NhZGFMaW5rLkNvbW11bmljYXRpb24u",
+            "R3JwY2IGcHJvdG8z"));
       descriptor = pbr::FileDescriptor.FromGeneratedCode(descriptorData,
           new pbr::FileDescriptor[] { global::Google.Protobuf.WellKnownTypes.TimestampReflection.Descriptor, global::Google.Protobuf.WellKnownTypes.WrappersReflection.Descriptor, },
           new pbr::GeneratedClrTypeInfo(new[] {typeof(global::ScadaLink.Communication.Grpc.Quality), typeof(global::ScadaLink.Communication.Grpc.AlarmStateEnum), typeof(global::ScadaLink.Communication.Grpc.AlarmLevelEnum), }, null, new pbr::GeneratedClrTypeInfo[] {
@@ -95,7 +101,9 @@ namespace ScadaLink.Communication.Grpc {
             new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.IngestAck), global::ScadaLink.Communication.Grpc.IngestAck.Parser, new[]{ "AcceptedEventIds" }, null, null, null, null),
             new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.SiteCallOperationalDto), global::ScadaLink.Communication.Grpc.SiteCallOperationalDto.Parser, new[]{ "TrackedOperationId", "Channel", "Target", "SourceSite", "Status", "RetryCount", "LastError", "HttpStatus", "CreatedAtUtc", "UpdatedAtUtc", "TerminalAtUtc" }, null, null, null, null),
             new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.CachedTelemetryPacket), global::ScadaLink.Communication.Grpc.CachedTelemetryPacket.Parser, new[]{ "AuditEvent", "Operational" }, null, null, null, null),
-            new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.CachedTelemetryBatch), global::ScadaLink.Communication.Grpc.CachedTelemetryBatch.Parser, new[]{ "Packets" }, null, null, null, null)
+            new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.CachedTelemetryBatch), global::ScadaLink.Communication.Grpc.CachedTelemetryBatch.Parser, new[]{ "Packets" }, null, null, null, null),
+            new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest), global::ScadaLink.Communication.Grpc.PullAuditEventsRequest.Parser, new[]{ "SinceUtc", "BatchSize" }, null, null, null, null),
+            new pbr::GeneratedClrTypeInfo(typeof(global::ScadaLink.Communication.Grpc.PullAuditEventsResponse), global::ScadaLink.Communication.Grpc.PullAuditEventsResponse.Parser, new[]{ "Events", "MoreAvailable" }, null, null, null, null)
           }));
     }
     #endregion
@@ -3862,6 +3870,482 @@ namespace ScadaLink.Communication.Grpc {
 
   }
 
+  /// <summary>
+  /// Audit Log (#23) M6 reconciliation pull: central→site request for any
+  /// site-local AuditLog rows with OccurredAtUtc >= since_utc that have not yet
+  /// been ingested centrally (ForwardState in {Pending, Forwarded}). The site
+  /// flips returned rows to Reconciled after the response is on the wire.
+  /// more_available signals batch_size was saturated so the caller knows to
+  /// issue a follow-up pull with an advanced since_utc cursor.
+  /// </summary>
+  [global::System.Diagnostics.DebuggerDisplayAttribute("{ToString(),nq}")]
+  public sealed partial class PullAuditEventsRequest : pb::IMessage<PullAuditEventsRequest>
+  #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      , pb::IBufferMessage
+  #endif
+  {
+    private static readonly pb::MessageParser<PullAuditEventsRequest> _parser = new pb::MessageParser<PullAuditEventsRequest>(() => new PullAuditEventsRequest());
+    private pb::UnknownFieldSet _unknownFields;
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public static pb::MessageParser<PullAuditEventsRequest> Parser { get { return _parser; } }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public static pbr::MessageDescriptor Descriptor {
+      get { return global::ScadaLink.Communication.Grpc.SitestreamReflection.Descriptor.MessageTypes[10]; }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    pbr::MessageDescriptor pb::IMessage.Descriptor {
+      get { return Descriptor; }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsRequest() {
+      OnConstruction();
+    }
+
+    partial void OnConstruction();
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsRequest(PullAuditEventsRequest other) : this() {
+      sinceUtc_ = other.sinceUtc_ != null ? other.sinceUtc_.Clone() : null;
+      batchSize_ = other.batchSize_;
+      _unknownFields = pb::UnknownFieldSet.Clone(other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsRequest Clone() {
+      return new PullAuditEventsRequest(this);
+    }
+
+    /// <summary>Field number for the "since_utc" field.</summary>
+    public const int SinceUtcFieldNumber = 1;
+    private global::Google.Protobuf.WellKnownTypes.Timestamp sinceUtc_;
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public global::Google.Protobuf.WellKnownTypes.Timestamp SinceUtc {
+      get { return sinceUtc_; }
+      set {
+        sinceUtc_ = value;
+      }
+    }
+
+    /// <summary>Field number for the "batch_size" field.</summary>
+    public const int BatchSizeFieldNumber = 2;
+    private int batchSize_;
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public int BatchSize {
+      get { return batchSize_; }
+      set {
+        batchSize_ = value;
+      }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override bool Equals(object other) {
+      return Equals(other as PullAuditEventsRequest);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public bool Equals(PullAuditEventsRequest other) {
+      if (ReferenceEquals(other, null)) {
+        return false;
+      }
+      if (ReferenceEquals(other, this)) {
+        return true;
+      }
+      if (!object.Equals(SinceUtc, other.SinceUtc)) return false;
+      if (BatchSize != other.BatchSize) return false;
+      return Equals(_unknownFields, other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override int GetHashCode() {
+      int hash = 1;
+      if (sinceUtc_ != null) hash ^= SinceUtc.GetHashCode();
+      if (BatchSize != 0) hash ^= BatchSize.GetHashCode();
+      if (_unknownFields != null) {
+        hash ^= _unknownFields.GetHashCode();
+      }
+      return hash;
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override string ToString() {
+      return pb::JsonFormatter.ToDiagnosticString(this);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void WriteTo(pb::CodedOutputStream output) {
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      output.WriteRawMessage(this);
+    #else
+      if (sinceUtc_ != null) {
+        output.WriteRawTag(10);
+        output.WriteMessage(SinceUtc);
+      }
+      if (BatchSize != 0) {
+        output.WriteRawTag(16);
+        output.WriteInt32(BatchSize);
+      }
+      if (_unknownFields != null) {
+        _unknownFields.WriteTo(output);
+      }
+    #endif
+    }
+
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    void pb::IBufferMessage.InternalWriteTo(ref pb::WriteContext output) {
+      if (sinceUtc_ != null) {
+        output.WriteRawTag(10);
+        output.WriteMessage(SinceUtc);
+      }
+      if (BatchSize != 0) {
+        output.WriteRawTag(16);
+        output.WriteInt32(BatchSize);
+      }
+      if (_unknownFields != null) {
+        _unknownFields.WriteTo(ref output);
+      }
+    }
+    #endif
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public int CalculateSize() {
+      int size = 0;
+      if (sinceUtc_ != null) {
+        size += 1 + pb::CodedOutputStream.ComputeMessageSize(SinceUtc);
+      }
+      if (BatchSize != 0) {
+        size += 1 + pb::CodedOutputStream.ComputeInt32Size(BatchSize);
+      }
+      if (_unknownFields != null) {
+        size += _unknownFields.CalculateSize();
+      }
+      return size;
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void MergeFrom(PullAuditEventsRequest other) {
+      if (other == null) {
+        return;
+      }
+      if (other.sinceUtc_ != null) {
+        if (sinceUtc_ == null) {
+          SinceUtc = new global::Google.Protobuf.WellKnownTypes.Timestamp();
+        }
+        SinceUtc.MergeFrom(other.SinceUtc);
+      }
+      if (other.BatchSize != 0) {
+        BatchSize = other.BatchSize;
+      }
+      _unknownFields = pb::UnknownFieldSet.MergeFrom(_unknownFields, other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void MergeFrom(pb::CodedInputStream input) {
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      input.ReadRawMessage(this);
+    #else
+      uint tag;
+      while ((tag = input.ReadTag()) != 0) {
+      if ((tag & 7) == 4) {
+        // Abort on any end group tag.
+        return;
+      }
+      switch(tag) {
+          default:
+            _unknownFields = pb::UnknownFieldSet.MergeFieldFrom(_unknownFields, input);
+            break;
+          case 10: {
+            if (sinceUtc_ == null) {
+              SinceUtc = new global::Google.Protobuf.WellKnownTypes.Timestamp();
+            }
+            input.ReadMessage(SinceUtc);
+            break;
+          }
+          case 16: {
+            BatchSize = input.ReadInt32();
+            break;
+          }
+        }
+      }
+    #endif
+    }
+
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    void pb::IBufferMessage.InternalMergeFrom(ref pb::ParseContext input) {
+      uint tag;
+      while ((tag = input.ReadTag()) != 0) {
+      if ((tag & 7) == 4) {
+        // Abort on any end group tag.
+        return;
+      }
+      switch(tag) {
+          default:
+            _unknownFields = pb::UnknownFieldSet.MergeFieldFrom(_unknownFields, ref input);
+            break;
+          case 10: {
+            if (sinceUtc_ == null) {
+              SinceUtc = new global::Google.Protobuf.WellKnownTypes.Timestamp();
+            }
+            input.ReadMessage(SinceUtc);
+            break;
+          }
+          case 16: {
+            BatchSize = input.ReadInt32();
+            break;
+          }
+        }
+      }
+    }
+    #endif
+
+  }
+
+  [global::System.Diagnostics.DebuggerDisplayAttribute("{ToString(),nq}")]
+  public sealed partial class PullAuditEventsResponse : pb::IMessage<PullAuditEventsResponse>
+  #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      , pb::IBufferMessage
+  #endif
+  {
+    private static readonly pb::MessageParser<PullAuditEventsResponse> _parser = new pb::MessageParser<PullAuditEventsResponse>(() => new PullAuditEventsResponse());
+    private pb::UnknownFieldSet _unknownFields;
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public static pb::MessageParser<PullAuditEventsResponse> Parser { get { return _parser; } }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public static pbr::MessageDescriptor Descriptor {
+      get { return global::ScadaLink.Communication.Grpc.SitestreamReflection.Descriptor.MessageTypes[11]; }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    pbr::MessageDescriptor pb::IMessage.Descriptor {
+      get { return Descriptor; }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsResponse() {
+      OnConstruction();
+    }
+
+    partial void OnConstruction();
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsResponse(PullAuditEventsResponse other) : this() {
+      events_ = other.events_.Clone();
+      moreAvailable_ = other.moreAvailable_;
+      _unknownFields = pb::UnknownFieldSet.Clone(other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public PullAuditEventsResponse Clone() {
+      return new PullAuditEventsResponse(this);
+    }
+
+    /// <summary>Field number for the "events" field.</summary>
+    public const int EventsFieldNumber = 1;
+    private static readonly pb::FieldCodec<global::ScadaLink.Communication.Grpc.AuditEventDto> _repeated_events_codec
+        = pb::FieldCodec.ForMessage(10, global::ScadaLink.Communication.Grpc.AuditEventDto.Parser);
+    private readonly pbc::RepeatedField<global::ScadaLink.Communication.Grpc.AuditEventDto> events_ = new pbc::RepeatedField<global::ScadaLink.Communication.Grpc.AuditEventDto>();
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public pbc::RepeatedField<global::ScadaLink.Communication.Grpc.AuditEventDto> Events {
+      get { return events_; }
+    }
+
+    /// <summary>Field number for the "more_available" field.</summary>
+    public const int MoreAvailableFieldNumber = 2;
+    private bool moreAvailable_;
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public bool MoreAvailable {
+      get { return moreAvailable_; }
+      set {
+        moreAvailable_ = value;
+      }
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override bool Equals(object other) {
+      return Equals(other as PullAuditEventsResponse);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public bool Equals(PullAuditEventsResponse other) {
+      if (ReferenceEquals(other, null)) {
+        return false;
+      }
+      if (ReferenceEquals(other, this)) {
+        return true;
+      }
+      if(!events_.Equals(other.events_)) return false;
+      if (MoreAvailable != other.MoreAvailable) return false;
+      return Equals(_unknownFields, other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override int GetHashCode() {
+      int hash = 1;
+      hash ^= events_.GetHashCode();
+      if (MoreAvailable != false) hash ^= MoreAvailable.GetHashCode();
+      if (_unknownFields != null) {
+        hash ^= _unknownFields.GetHashCode();
+      }
+      return hash;
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public override string ToString() {
+      return pb::JsonFormatter.ToDiagnosticString(this);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void WriteTo(pb::CodedOutputStream output) {
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      output.WriteRawMessage(this);
+    #else
+      events_.WriteTo(output, _repeated_events_codec);
+      if (MoreAvailable != false) {
+        output.WriteRawTag(16);
+        output.WriteBool(MoreAvailable);
+      }
+      if (_unknownFields != null) {
+        _unknownFields.WriteTo(output);
+      }
+    #endif
+    }
+
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    void pb::IBufferMessage.InternalWriteTo(ref pb::WriteContext output) {
+      events_.WriteTo(ref output, _repeated_events_codec);
+      if (MoreAvailable != false) {
+        output.WriteRawTag(16);
+        output.WriteBool(MoreAvailable);
+      }
+      if (_unknownFields != null) {
+        _unknownFields.WriteTo(ref output);
+      }
+    }
+    #endif
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public int CalculateSize() {
+      int size = 0;
+      size += events_.CalculateSize(_repeated_events_codec);
+      if (MoreAvailable != false) {
+        size += 1 + 1;
+      }
+      if (_unknownFields != null) {
+        size += _unknownFields.CalculateSize();
+      }
+      return size;
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void MergeFrom(PullAuditEventsResponse other) {
+      if (other == null) {
+        return;
+      }
+      events_.Add(other.events_);
+      if (other.MoreAvailable != false) {
+        MoreAvailable = other.MoreAvailable;
+      }
+      _unknownFields = pb::UnknownFieldSet.MergeFrom(_unknownFields, other._unknownFields);
+    }
+
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    public void MergeFrom(pb::CodedInputStream input) {
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+      input.ReadRawMessage(this);
+    #else
+      uint tag;
+      while ((tag = input.ReadTag()) != 0) {
+      if ((tag & 7) == 4) {
+        // Abort on any end group tag.
+        return;
+      }
+      switch(tag) {
+          default:
+            _unknownFields = pb::UnknownFieldSet.MergeFieldFrom(_unknownFields, input);
+            break;
+          case 10: {
+            events_.AddEntriesFrom(input, _repeated_events_codec);
+            break;
+          }
+          case 16: {
+            MoreAvailable = input.ReadBool();
+            break;
+          }
+        }
+      }
+    #endif
+    }
+
+    #if !GOOGLE_PROTOBUF_REFSTRUCT_COMPATIBILITY_MODE
+    [global::System.Diagnostics.DebuggerNonUserCodeAttribute]
+    [global::System.CodeDom.Compiler.GeneratedCode("protoc", null)]
+    void pb::IBufferMessage.InternalMergeFrom(ref pb::ParseContext input) {
+      uint tag;
+      while ((tag = input.ReadTag()) != 0) {
+      if ((tag & 7) == 4) {
+        // Abort on any end group tag.
+        return;
+      }
+      switch(tag) {
+          default:
+            _unknownFields = pb::UnknownFieldSet.MergeFieldFrom(_unknownFields, ref input);
+            break;
+          case 10: {
+            events_.AddEntriesFrom(ref input, _repeated_events_codec);
+            break;
+          }
+          case 16: {
+            MoreAvailable = input.ReadBool();
+            break;
+          }
+        }
+      }
+    }
+    #endif
+
+  }
+
   #endregion
 
 }

src/ScadaLink.Communication/SiteStreamGrpc/SitestreamGrpc.cs

@@ -55,6 +55,10 @@ namespace ScadaLink.Communication.Grpc {
     static readonly grpc::Marshaller<global::ScadaLink.Communication.Grpc.IngestAck> __Marshaller_sitestream_IngestAck = grpc::Marshallers.Create(__Helper_SerializeMessage, context => __Helper_DeserializeMessage(context, global::ScadaLink.Communication.Grpc.IngestAck.Parser));
     [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
     static readonly grpc::Marshaller<global::ScadaLink.Communication.Grpc.CachedTelemetryBatch> __Marshaller_sitestream_CachedTelemetryBatch = grpc::Marshallers.Create(__Helper_SerializeMessage, context => __Helper_DeserializeMessage(context, global::ScadaLink.Communication.Grpc.CachedTelemetryBatch.Parser));
+    [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+    static readonly grpc::Marshaller<global::ScadaLink.Communication.Grpc.PullAuditEventsRequest> __Marshaller_sitestream_PullAuditEventsRequest = grpc::Marshallers.Create(__Helper_SerializeMessage, context => __Helper_DeserializeMessage(context, global::ScadaLink.Communication.Grpc.PullAuditEventsRequest.Parser));
+    [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+    static readonly grpc::Marshaller<global::ScadaLink.Communication.Grpc.PullAuditEventsResponse> __Marshaller_sitestream_PullAuditEventsResponse = grpc::Marshallers.Create(__Helper_SerializeMessage, context => __Helper_DeserializeMessage(context, global::ScadaLink.Communication.Grpc.PullAuditEventsResponse.Parser));
 
     [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
     static readonly grpc::Method<global::ScadaLink.Communication.Grpc.InstanceStreamRequest, global::ScadaLink.Communication.Grpc.SiteStreamEvent> __Method_SubscribeInstance = new grpc::Method<global::ScadaLink.Communication.Grpc.InstanceStreamRequest, global::ScadaLink.Communication.Grpc.SiteStreamEvent>(
@@ -80,6 +84,14 @@ namespace ScadaLink.Communication.Grpc {
         __Marshaller_sitestream_CachedTelemetryBatch,
         __Marshaller_sitestream_IngestAck);
 
+    [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+    static readonly grpc::Method<global::ScadaLink.Communication.Grpc.PullAuditEventsRequest, global::ScadaLink.Communication.Grpc.PullAuditEventsResponse> __Method_PullAuditEvents = new grpc::Method<global::ScadaLink.Communication.Grpc.PullAuditEventsRequest, global::ScadaLink.Communication.Grpc.PullAuditEventsResponse>(
+        grpc::MethodType.Unary,
+        __ServiceName,
+        "PullAuditEvents",
+        __Marshaller_sitestream_PullAuditEventsRequest,
+        __Marshaller_sitestream_PullAuditEventsResponse);
+
     /// <summary>Service descriptor</summary>
     public static global::Google.Protobuf.Reflection.ServiceDescriptor Descriptor
     {
@@ -108,6 +120,12 @@ namespace ScadaLink.Communication.Grpc {
         throw new grpc::RpcException(new grpc::Status(grpc::StatusCode.Unimplemented, ""));
       }
 
+      [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+      public virtual global::System.Threading.Tasks.Task<global::ScadaLink.Communication.Grpc.PullAuditEventsResponse> PullAuditEvents(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest request, grpc::ServerCallContext context)
+      {
+        throw new grpc::RpcException(new grpc::Status(grpc::StatusCode.Unimplemented, ""));
+      }
+
     }
 
     /// <summary>Client for SiteStreamService</summary>
@@ -187,6 +205,26 @@ namespace ScadaLink.Communication.Grpc {
       {
         return CallInvoker.AsyncUnaryCall(__Method_IngestCachedTelemetry, null, options, request);
       }
+      [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+      public virtual global::ScadaLink.Communication.Grpc.PullAuditEventsResponse PullAuditEvents(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest request, grpc::Metadata headers = null, global::System.DateTime? deadline = null, global::System.Threading.CancellationToken cancellationToken = default(global::System.Threading.CancellationToken))
+      {
+        return PullAuditEvents(request, new grpc::CallOptions(headers, deadline, cancellationToken));
+      }
+      [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+      public virtual global::ScadaLink.Communication.Grpc.PullAuditEventsResponse PullAuditEvents(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest request, grpc::CallOptions options)
+      {
+        return CallInvoker.BlockingUnaryCall(__Method_PullAuditEvents, null, options, request);
+      }
+      [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+      public virtual grpc::AsyncUnaryCall<global::ScadaLink.Communication.Grpc.PullAuditEventsResponse> PullAuditEventsAsync(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest request, grpc::Metadata headers = null, global::System.DateTime? deadline = null, global::System.Threading.CancellationToken cancellationToken = default(global::System.Threading.CancellationToken))
+      {
+        return PullAuditEventsAsync(request, new grpc::CallOptions(headers, deadline, cancellationToken));
+      }
+      [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
+      public virtual grpc::AsyncUnaryCall<global::ScadaLink.Communication.Grpc.PullAuditEventsResponse> PullAuditEventsAsync(global::ScadaLink.Communication.Grpc.PullAuditEventsRequest request, grpc::CallOptions options)
+      {
+        return CallInvoker.AsyncUnaryCall(__Method_PullAuditEvents, null, options, request);
+      }
       /// <summary>Creates a new instance of client from given <c>ClientBaseConfiguration</c>.</summary>
       [global::System.CodeDom.Compiler.GeneratedCode("grpc_csharp_plugin", null)]
       protected override SiteStreamServiceClient NewInstance(ClientBaseConfiguration configuration)
@@ -203,7 +241,8 @@ namespace ScadaLink.Communication.Grpc {
       return grpc::ServerServiceDefinition.CreateBuilder()
           .AddMethod(__Method_SubscribeInstance, serviceImpl.SubscribeInstance)
           .AddMethod(__Method_IngestAuditEvents, serviceImpl.IngestAuditEvents)
-          .AddMethod(__Method_IngestCachedTelemetry, serviceImpl.IngestCachedTelemetry).Build();
+          .AddMethod(__Method_IngestCachedTelemetry, serviceImpl.IngestCachedTelemetry)
+          .AddMethod(__Method_PullAuditEvents, serviceImpl.PullAuditEvents).Build();
     }
 
     /// <summary>Register service method with a service binder with or without implementation. Useful when customizing the service binding logic.
@@ -216,6 +255,7 @@ namespace ScadaLink.Communication.Grpc {
       serviceBinder.AddMethod(__Method_SubscribeInstance, serviceImpl == null ? null : new grpc::ServerStreamingServerMethod<global::ScadaLink.Communication.Grpc.InstanceStreamRequest, global::ScadaLink.Communication.Grpc.SiteStreamEvent>(serviceImpl.SubscribeInstance));
       serviceBinder.AddMethod(__Method_IngestAuditEvents, serviceImpl == null ? null : new grpc::UnaryServerMethod<global::ScadaLink.Communication.Grpc.AuditEventBatch, global::ScadaLink.Communication.Grpc.IngestAck>(serviceImpl.IngestAuditEvents));
       serviceBinder.AddMethod(__Method_IngestCachedTelemetry, serviceImpl == null ? null : new grpc::UnaryServerMethod<global::ScadaLink.Communication.Grpc.CachedTelemetryBatch, global::ScadaLink.Communication.Grpc.IngestAck>(serviceImpl.IngestCachedTelemetry));
+      serviceBinder.AddMethod(__Method_PullAuditEvents, serviceImpl == null ? null : new grpc::UnaryServerMethod<global::ScadaLink.Communication.Grpc.PullAuditEventsRequest, global::ScadaLink.Communication.Grpc.PullAuditEventsResponse>(serviceImpl.PullAuditEvents));
     }
 
   }

src/ScadaLink.ConfigurationDatabase/Maintenance/AuditLogPartitionMaintenance.cs

@@ -0,0 +1,218 @@
+using System.Globalization;
+using Microsoft.Data.SqlClient;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Logging.Abstractions;
+using ScadaLink.Commons.Interfaces;
+
+namespace ScadaLink.ConfigurationDatabase.Maintenance;
+
+/// <summary>
+/// EF/SQL-Server implementation of <see cref="IPartitionMaintenance"/> that
+/// rolls forward <c>pf_AuditLog_Month</c> by issuing
+/// <c>ALTER PARTITION FUNCTION … SPLIT RANGE</c> for each missing future
+/// monthly boundary.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The class is scoped (registered alongside the other repositories in
+/// <c>AddConfigurationDatabase</c>) because it shares <see cref="ScadaLinkDbContext"/>
+/// — the hosted service opens a per-tick DI scope, resolves a fresh instance,
+/// and lets the scope's <c>DbContext</c> dispose with it. The class itself
+/// holds no state between calls.
+/// </para>
+/// <para>
+/// <b>Idempotency model.</b> Each tick reads the current max boundary from
+/// <c>sys.partition_range_values</c> and only issues SPLIT RANGE for
+/// boundaries that strictly follow it — a boundary already covered is never
+/// re-issued, so the "boundary already exists" failure (SQL Server msg 7708
+/// / 7711) is avoided by construction rather than caught. The pre-check is
+/// cheaper than the alternative TRY/CATCH around every SPLIT call and also
+/// keeps the returned <c>added</c> list semantically precise.
+/// </para>
+/// <para>
+/// <b>Why "first of next month".</b> The migration seeds boundaries on the
+/// first-of-month at midnight UTC; we preserve that convention so the
+/// resulting partition layout is uniform. <see cref="NormalizeToFirstOfMonth"/>
+/// rounds an arbitrary timestamp up to the next first-of-month boundary
+/// (e.g. 2026-05-20 → 2026-06-01), and <see cref="NextMonthBoundary"/>
+/// walks one month at a time from there.
+/// </para>
+/// <para>
+/// <b>Permissions.</b> The migration's <c>scadalink_audit_purger</c> role
+/// already carries <c>ALTER ON SCHEMA::dbo</c>, which is sufficient for
+/// <c>ALTER PARTITION FUNCTION SPLIT RANGE</c>. No additional grant is
+/// required.
+/// </para>
+/// </remarks>
+public sealed class AuditLogPartitionMaintenance : IPartitionMaintenance
+{
+    private const string PartitionFunctionName = "pf_AuditLog_Month";
+    private const string PartitionSchemeName = "ps_AuditLog_Month";
+    private const string TargetFileGroup = "PRIMARY";
+
+    private readonly ScadaLinkDbContext _context;
+    private readonly ILogger<AuditLogPartitionMaintenance> _logger;
+
+    public AuditLogPartitionMaintenance(
+        ScadaLinkDbContext context,
+        ILogger<AuditLogPartitionMaintenance>? logger = null)
+    {
+        _context = context ?? throw new ArgumentNullException(nameof(context));
+        _logger = logger ?? NullLogger<AuditLogPartitionMaintenance>.Instance;
+    }
+
+    /// <inheritdoc />
+    public async Task<DateTime?> GetMaxBoundaryAsync(CancellationToken ct = default)
+    {
+        // CAST the sql_variant `value` column to datetime2(7) — every boundary in
+        // pf_AuditLog_Month is declared as datetime2(7) by the migration, so the
+        // cast never loses precision.
+        const string sql = @"
+SELECT MAX(CAST(rv.value AS datetime2(7)))
+FROM sys.partition_range_values rv
+INNER JOIN sys.partition_functions pf ON rv.function_id = pf.function_id
+WHERE pf.name = 'pf_AuditLog_Month';";
+
+        var conn = _context.Database.GetDbConnection();
+        var openedHere = false;
+        if (conn.State != System.Data.ConnectionState.Open)
+        {
+            await conn.OpenAsync(ct).ConfigureAwait(false);
+            openedHere = true;
+        }
+
+        try
+        {
+            await using var cmd = conn.CreateCommand();
+            cmd.CommandText = sql;
+            var raw = await cmd.ExecuteScalarAsync(ct).ConfigureAwait(false);
+            if (raw is null || raw is DBNull)
+            {
+                return null;
+            }
+
+            // ExecuteScalarAsync materialises datetime2 as DateTime with
+            // DateTimeKind.Unspecified; the boundary values are stored at
+            // UTC midnight by convention (migration seeds with 'T00:00:00'),
+            // so we re-tag the kind so downstream comparisons against
+            // DateTime.UtcNow stay in the same kind space.
+            var dt = (DateTime)raw;
+            return DateTime.SpecifyKind(dt, DateTimeKind.Utc);
+        }
+        finally
+        {
+            if (openedHere)
+            {
+                await conn.CloseAsync().ConfigureAwait(false);
+            }
+        }
+    }
+
+    /// <inheritdoc />
+    public async Task<IReadOnlyList<DateTime>> EnsureLookaheadAsync(
+        int lookaheadMonths,
+        CancellationToken ct = default)
+    {
+        if (lookaheadMonths < 1)
+        {
+            throw new ArgumentOutOfRangeException(
+                nameof(lookaheadMonths),
+                lookaheadMonths,
+                "Lookahead must be at least one month — the partition function would otherwise be allowed to fall behind 'now'.");
+        }
+
+        var nowUtc = DateTime.UtcNow;
+        // Horizon: the FIRST-OF-MONTH that must be the strictly-greater-than
+        // max boundary after this call. Example: nowUtc = 2026-05-20 and
+        // lookaheadMonths = 1 → horizon = 2026-07-01 (so the partition for
+        // June 2026 is already in place by mid-May).
+        var horizon = NormalizeToFirstOfMonth(nowUtc).AddMonths(lookaheadMonths);
+
+        var max = await GetMaxBoundaryAsync(ct).ConfigureAwait(false);
+        if (max is null)
+        {
+            // No partition function (e.g. migrations not applied) — nothing
+            // we can safely SPLIT against. Log and return; the absence is a
+            // genuine misconfiguration that other parts of the system will
+            // surface louder than we could here.
+            _logger.LogWarning(
+                "EnsureLookaheadAsync: partition function {PartitionFunctionName} not found; skipping.",
+                PartitionFunctionName);
+            return Array.Empty<DateTime>();
+        }
+
+        // Start splitting from the FIRST month strictly after max — if max is
+        // already first-of-month (the common case), that's max + 1 month;
+        // otherwise NormalizeToFirstOfMonth rounds up.
+        var next = NormalizeToFirstOfMonth(max.Value.AddDays(1));
+
+        // Edge case: max already past horizon → no work to do.
+        if (next > horizon)
+        {
+            return Array.Empty<DateTime>();
+        }
+
+        var added = new List<DateTime>();
+        while (next <= horizon)
+        {
+            // Boundary literal must be a deterministic, culture-invariant ISO
+            // string — SQL Server parses it as datetime2 via implicit conversion.
+            // SPLIT RANGE does NOT accept @-parameters; the value is part of the
+            // DDL statement, so we render it directly. The format is
+            // guaranteed (yyyy-MM-ddTHH:mm:ss.fffffff) so there is no injection
+            // surface.
+            var literal = next.ToString("yyyy-MM-ddTHH:mm:ss.fffffff", CultureInfo.InvariantCulture);
+
+            // Before every SPLIT we must (re-)set the NEXT USED filegroup on
+            // ps_AuditLog_Month. Even though the scheme was created with
+            // `ALL TO ([PRIMARY])` (which auto-populates NEXT USED once), SQL
+            // Server consumes that hint on the FIRST split — subsequent splits
+            // raise msg 7707 ("partition scheme … does not have any next used
+            // filegroup") unless NEXT USED is explicitly re-set. Re-issuing it
+            // before every split is idempotent and keeps the loop simple.
+            var sql = $@"
+ALTER PARTITION SCHEME {PartitionSchemeName} NEXT USED [{TargetFileGroup}];
+ALTER PARTITION FUNCTION {PartitionFunctionName}() SPLIT RANGE ('{literal}');";
+
+            try
+            {
+                await _context.Database.ExecuteSqlRawAsync(sql, ct).ConfigureAwait(false);
+                added.Add(next);
+            }
+            catch (SqlException ex)
+            {
+                // Belt-and-braces: even though we read max-boundary first, an
+                // ALTER from another process could have raced us. Logging at
+                // Warning rather than Error because the desired end state
+                // (boundary present) is satisfied by either path.
+                _logger.LogWarning(
+                    ex,
+                    "EnsureLookaheadAsync: SPLIT RANGE for boundary {Boundary:o} failed; continuing.",
+                    next);
+            }
+
+            next = NextMonthBoundary(next);
+        }
+
+        return added;
+    }
+
+    /// <summary>
+    /// Rounds an arbitrary instant UP to the next first-of-month UTC. Inputs
+    /// that ARE already a first-of-month at midnight are returned as-is so
+    /// callers can compose this freely without double-incrementing.
+    /// </summary>
+    private static DateTime NormalizeToFirstOfMonth(DateTime instant)
+    {
+        var utc = instant.Kind == DateTimeKind.Utc
+            ? instant
+            : DateTime.SpecifyKind(instant, DateTimeKind.Utc);
+
+        var firstOfThisMonth = new DateTime(utc.Year, utc.Month, 1, 0, 0, 0, DateTimeKind.Utc);
+        return utc == firstOfThisMonth ? firstOfThisMonth : firstOfThisMonth.AddMonths(1);
+    }
+
+    private static DateTime NextMonthBoundary(DateTime boundary) =>
+        boundary.AddMonths(1);
+}

src/ScadaLink.ConfigurationDatabase/Repositories/AuditLogRepository.cs

@@ -179,18 +179,246 @@ VALUES
     }
 
     /// <summary>
-    /// M1 honest contract: throws <see cref="NotSupportedException"/>. The
-    /// <c>UX_AuditLog_EventId</c> unique index is non-aligned with
-    /// <c>ps_AuditLog_Month</c> (it lives on <c>[PRIMARY]</c> to keep
-    /// <see cref="InsertIfNotExistsAsync"/> cheap), and SQL Server rejects
-    /// <c>ALTER TABLE … SWITCH PARTITION</c> when a non-aligned index is present.
-    /// The drop-and-rebuild dance that makes the switch legal ships with the M6
-    /// purge actor.
+    /// M6-T4 production implementation of the drop-and-rebuild dance documented
+    /// on <see cref="IAuditLogRepository.SwitchOutPartitionAsync"/>.
     /// </summary>
-    public Task SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default)
+    /// <remarks>
+    /// <para>
+    /// The staging table name is GUID-suffixed so concurrent purge attempts on
+    /// different boundaries cannot collide. The staging schema is byte-identical
+    /// to the live <c>AuditLog</c> table (same column types, lengths,
+    /// nullability, and clustered-key shape) — SQL Server's
+    /// <c>ALTER TABLE … SWITCH PARTITION</c> rejects any drift. Keep this CREATE
+    /// in sync with both the migration that ships the live table
+    /// (<c>20260520142214_AddAuditLogTable</c>) and
+    /// <c>AuditLogEntityTypeConfiguration</c>.
+    /// </para>
+    /// <para>
+    /// All five steps run inside an explicit transaction so the SWITCH +
+    /// staging-DROP are atomic from the perspective of a consumer reading via
+    /// snapshot isolation; the CATCH rolls back and runs an idempotent
+    /// "rebuild UX_AuditLog_EventId if it doesn't exist" so a partial failure
+    /// never leaves the live table without its idempotency-supporting unique
+    /// index.
+    /// </para>
+    /// </remarks>
+    public async Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default)
     {
-        throw new NotSupportedException(
-            "AuditLog partition switch is blocked by the non-aligned UX_AuditLog_EventId " +
-            "unique index; the drop-and-rebuild dance ships in M6 (purge actor).");
+        // GUID-suffixed staging name: prevents collision with any concurrent
+        // purge attempt and avoids polluting the AuditLog object namespace with
+        // a predictable identifier.
+        var stagingTableName = $"AuditLog_Staging_{Guid.NewGuid():N}";
+
+        // ISO 8601 in UTC — SQL Server's datetime2 literal parser accepts this
+        // unambiguously and the value is round-trip-safe across SET DATEFORMAT
+        // settings.
+        var monthBoundaryStr = monthBoundary.ToUniversalTime().ToString("yyyy-MM-dd HH:mm:ss");
+
+        // Two-statement batch: the first SELECT samples the per-partition row
+        // count BEFORE the dance so we can report it back to the purge actor;
+        // the second batch performs the drop-and-rebuild. We use OUTPUT-style
+        // variables wired through @@ROWCOUNT after the SWITCH is not viable
+        // because SWITCH is a metadata-only operation that doesn't move rows in
+        // a way @@ROWCOUNT can observe.
+        var sampleSql = $@"
+            SELECT COUNT_BIG(*) FROM dbo.AuditLog
+            WHERE $PARTITION.pf_AuditLog_Month(OccurredAtUtc) =
+                  $partition.pf_AuditLog_Month('{monthBoundaryStr}');";
+
+        var sql = $@"
+        BEGIN TRY
+            BEGIN TRANSACTION;
+
+            -- 1. Drop the non-aligned unique index. ALTER TABLE SWITCH refuses
+            --    to run while it exists.
+            IF EXISTS (SELECT 1 FROM sys.indexes WHERE name = 'UX_AuditLog_EventId' AND object_id = OBJECT_ID('dbo.AuditLog'))
+                DROP INDEX UX_AuditLog_EventId ON dbo.AuditLog;
+
+            -- 2. Staging table on [PRIMARY] (non-partitioned) with column shapes
+            --    byte-identical to dbo.AuditLog. Any drift here causes SWITCH to
+            --    reject the operation with msg 4904/4915.
+            CREATE TABLE dbo.[{stagingTableName}] (
+                EventId            uniqueidentifier NOT NULL,
+                OccurredAtUtc      datetime2(7)     NOT NULL,
+                IngestedAtUtc      datetime2(7)     NULL,
+                Channel            varchar(32)      NOT NULL,
+                Kind               varchar(32)      NOT NULL,
+                CorrelationId      uniqueidentifier NULL,
+                SourceSiteId       varchar(64)      NULL,
+                SourceInstanceId   varchar(128)     NULL,
+                SourceScript       varchar(128)     NULL,
+                Actor              varchar(128)     NULL,
+                Target             varchar(256)     NULL,
+                Status             varchar(32)      NOT NULL,
+                HttpStatus         int              NULL,
+                DurationMs         int              NULL,
+                ErrorMessage       nvarchar(1024)   NULL,
+                ErrorDetail        nvarchar(max)    NULL,
+                RequestSummary     nvarchar(max)    NULL,
+                ResponseSummary    nvarchar(max)    NULL,
+                PayloadTruncated   bit              NOT NULL,
+                Extra              nvarchar(max)    NULL,
+                ForwardState       varchar(32)      NULL,
+                CONSTRAINT PK_{stagingTableName} PRIMARY KEY CLUSTERED (EventId, OccurredAtUtc)
+            ) ON [PRIMARY];
+
+            -- 3. Switch the partition out. $partition.pf_AuditLog_Month returns
+            --    the partition number that contains the supplied boundary value;
+            --    SWITCH PARTITION N moves that partition's pages to the staging
+            --    table (metadata-only, no row copying).
+            DECLARE @partitionNumber int = $partition.pf_AuditLog_Month('{monthBoundaryStr}');
+            DECLARE @sql nvarchar(max) = 'ALTER TABLE dbo.AuditLog SWITCH PARTITION ' + CAST(@partitionNumber AS nvarchar(10)) + ' TO dbo.[{stagingTableName}];';
+            EXEC sp_executesql @sql;
+
+            -- 4. Drop staging — the rows are discarded here. This is the purge.
+            DROP TABLE dbo.[{stagingTableName}];
+
+            -- 5. Rebuild the non-aligned unique index. Live traffic that hit the
+            --    table during steps 1-4 saw composite-PK uniqueness only; from
+            --    here on, single-column EventId uniqueness is restored.
+            CREATE UNIQUE NONCLUSTERED INDEX UX_AuditLog_EventId ON dbo.AuditLog (EventId) ON [PRIMARY];
+
+            COMMIT TRANSACTION;
+        END TRY
+        BEGIN CATCH
+            IF @@TRANCOUNT > 0 ROLLBACK TRANSACTION;
+
+            -- Best-effort staging cleanup. The DROP INDEX in step 1 is now
+            -- rolled back (so the index is back), but the staging table from
+            -- step 2 may or may not survive the rollback depending on the
+            -- failure point. Guard the DROP so a missing staging table doesn't
+            -- mask the original error.
+            IF OBJECT_ID('dbo.[{stagingTableName}]', 'U') IS NOT NULL DROP TABLE dbo.[{stagingTableName}];
+
+            -- Idempotent index rebuild — covers the niche case where ROLLBACK
+            -- failed to restore UX_AuditLog_EventId (or the failure happened
+            -- AFTER the COMMIT, which shouldn't be possible inside this TRY
+            -- but is cheap insurance). Without this, a failed switch could
+            -- leave the live table without its idempotency-supporting index.
+            IF NOT EXISTS (SELECT 1 FROM sys.indexes WHERE name = 'UX_AuditLog_EventId' AND object_id = OBJECT_ID('dbo.AuditLog'))
+                CREATE UNIQUE NONCLUSTERED INDEX UX_AuditLog_EventId ON dbo.AuditLog (EventId) ON [PRIMARY];
+
+            -- Surface the original error to the caller — the purge actor logs
+            -- and continues with the next boundary.
+            THROW;
+        END CATCH;";
+
+        // Sample the row count before the switch. The sample is best-effort
+        // (no transaction wrapping the sample-then-switch pair) because the
+        // central singleton is the only writer to this RPC and a daily-purge
+        // tick doesn't compete with concurrent SwitchOut callers. A
+        // concurrent INSERT racing the sample under-reports by at most a
+        // few rows, which is acceptable for an "approximate" purged-row
+        // count surfaced via AuditLogPurgedEvent.
+        long rowsDeleted = 0;
+        var conn = _context.Database.GetDbConnection();
+        var openedHere = false;
+        if (conn.State != System.Data.ConnectionState.Open)
+        {
+            await conn.OpenAsync(ct).ConfigureAwait(false);
+            openedHere = true;
+        }
+        try
+        {
+            await using (var sampleCmd = conn.CreateCommand())
+            {
+                sampleCmd.CommandText = sampleSql;
+                var sampleResult = await sampleCmd.ExecuteScalarAsync(ct).ConfigureAwait(false);
+                if (sampleResult is not null && sampleResult is not DBNull)
+                {
+                    rowsDeleted = Convert.ToInt64(sampleResult);
+                }
+            }
+        }
+        finally
+        {
+            if (openedHere)
+            {
+                await conn.CloseAsync().ConfigureAwait(false);
+            }
+        }
+
+        await _context.Database.ExecuteSqlRawAsync(sql, ct);
+        return rowsDeleted;
+    }
+
+    /// <summary>
+    /// Returns the set of <c>pf_AuditLog_Month</c> boundaries whose partition's
+    /// <c>MAX(OccurredAtUtc)</c> is strictly older than <paramref name="threshold"/>.
+    /// Boundaries with empty partitions are excluded — purging an empty
+    /// partition is wasted I/O.
+    /// </summary>
+    /// <remarks>
+    /// <para>
+    /// The CTE pulls every boundary value defined by the partition function and
+    /// joins it (via <c>$PARTITION.pf_AuditLog_Month</c>) to the live AuditLog
+    /// to compute per-partition <c>MAX(OccurredAtUtc)</c>. The outer filter
+    /// keeps only those whose MAX is non-NULL (partition has rows) AND strictly
+    /// less than the threshold (every row is past retention).
+    /// </para>
+    /// <para>
+    /// Note: the query scans the live <c>OccurredAtUtc</c> column to compute
+    /// the MAX per partition. With <c>IX_AuditLog_OccurredAtUtc</c> on the
+    /// partition-aligned scheme this is a single index seek per partition; for
+    /// 24 partitions and a daily purge cadence the cost is negligible.
+    /// </para>
+    /// </remarks>
+    public async Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+        DateTime threshold,
+        CancellationToken ct = default)
+    {
+        var thresholdUtc = threshold.ToUniversalTime();
+        var thresholdStr = thresholdUtc.ToString("yyyy-MM-dd HH:mm:ss.fffffff");
+
+        // Per-partition MAX over the live table. We materialise the boundary
+        // list first (24 rows) then LEFT JOIN to the MAX aggregate so empty
+        // partitions surface as NULL and get filtered out by the WHERE clause.
+        var sql = $@"
+            WITH Boundaries AS (
+                SELECT CAST(rv.value AS datetime2(7)) AS BoundaryValue,
+                       rv.boundary_id AS BoundaryId
+                FROM sys.partition_range_values rv
+                INNER JOIN sys.partition_functions pf ON rv.function_id = pf.function_id
+                WHERE pf.name = 'pf_AuditLog_Month'
+            )
+            SELECT b.BoundaryValue
+            FROM Boundaries b
+            CROSS APPLY (
+                SELECT MAX(a.OccurredAtUtc) AS MaxOccurredAt
+                FROM dbo.AuditLog a
+                WHERE $PARTITION.pf_AuditLog_Month(a.OccurredAtUtc) = b.BoundaryId + 1
+            ) x
+            WHERE x.MaxOccurredAt IS NOT NULL
+              AND x.MaxOccurredAt < CAST('{thresholdStr}' AS datetime2(7))
+            ORDER BY b.BoundaryValue;";
+
+        var conn = _context.Database.GetDbConnection();
+        var openedHere = false;
+        if (conn.State != System.Data.ConnectionState.Open)
+        {
+            await conn.OpenAsync(ct).ConfigureAwait(false);
+            openedHere = true;
+        }
+
+        var results = new List<DateTime>();
+        try
+        {
+            await using var cmd = conn.CreateCommand();
+            cmd.CommandText = sql;
+            await using var reader = await cmd.ExecuteReaderAsync(ct).ConfigureAwait(false);
+            while (await reader.ReadAsync(ct).ConfigureAwait(false))
+            {
+                results.Add(reader.GetDateTime(0));
+            }
+        }
+        finally
+        {
+            if (openedHere)
+            {
+                await conn.CloseAsync().ConfigureAwait(false);
+            }
+        }
+
+        return results;
     }
 }

src/ScadaLink.ConfigurationDatabase/ServiceCollectionExtensions.cs

@@ -1,8 +1,10 @@
 using Microsoft.AspNetCore.DataProtection;
 using Microsoft.EntityFrameworkCore;
 using Microsoft.Extensions.DependencyInjection;
+using ScadaLink.Commons.Interfaces;
 using ScadaLink.Commons.Interfaces.Repositories;
 using ScadaLink.Commons.Interfaces.Services;
+using ScadaLink.ConfigurationDatabase.Maintenance;
 using ScadaLink.ConfigurationDatabase.Repositories;
 using ScadaLink.ConfigurationDatabase.Services;
 
@@ -52,6 +54,13 @@ public static class ServiceCollectionExtensions
         services.AddScoped<IAuditService, AuditService>();
         services.AddScoped<IInstanceLocator, InstanceLocator>();
 
+        // #23 M6 Bundle D: IPartitionMaintenance drives the daily roll-forward
+        // of pf_AuditLog_Month from the central AuditLogPartitionMaintenanceService
+        // hosted service. Scoped because the implementation reuses the per-scope
+        // ScadaLinkDbContext for raw-SQL execution; the hosted service opens a
+        // fresh scope on each tick (mirrors AuditLogPurgeActor / AuditLogIngestActor).
+        services.AddScoped<IPartitionMaintenance, AuditLogPartitionMaintenance>();
+
         services.AddDataProtection()
             .PersistKeysToDbContext<ScadaLinkDbContext>();
 

src/ScadaLink.HealthMonitoring/ISiteHealthCollector.cs

@@ -1,4 +1,5 @@
 using ScadaLink.Commons.Messages.Health;
+using ScadaLink.Commons.Types;
 using ScadaLink.Commons.Types.Enums;
 
 namespace ScadaLink.HealthMonitoring;
@@ -28,6 +29,15 @@ public interface ISiteHealthCollector
     /// <c>AddAuditLogHealthMetricsBridge()</c>.
     /// </summary>
     void IncrementAuditRedactionFailure();
+    /// <summary>
+    /// Audit Log (#23) M6 Bundle E (T6) — replace the latest site-local
+    /// audit-queue backlog snapshot (pending count, oldest pending row,
+    /// on-disk file bytes) used by the next <see cref="CollectReport"/> call.
+    /// Refreshed periodically by the <c>SiteAuditBacklogReporter</c> hosted
+    /// service so each report carries a recent point-in-time view of the
+    /// site→central drain health.
+    /// </summary>
+    void UpdateSiteAuditBacklog(SiteAuditBacklogSnapshot snapshot);
     void UpdateConnectionHealth(string connectionName, ConnectionHealth health);
     void RemoveConnection(string connectionName);
     void UpdateTagResolution(string connectionName, int totalSubscribed, int successfullyResolved);

src/ScadaLink.HealthMonitoring/SiteHealthCollector.cs

@@ -1,5 +1,6 @@
 using System.Collections.Concurrent;
 using ScadaLink.Commons.Messages.Health;
+using ScadaLink.Commons.Types;
 using ScadaLink.Commons.Types.Enums;
 
 namespace ScadaLink.HealthMonitoring;
@@ -15,6 +16,7 @@ public class SiteHealthCollector : ISiteHealthCollector
     private int _deadLetterCount;
     private int _siteAuditWriteFailures;
     private int _auditRedactionFailures;
+    private volatile SiteAuditBacklogSnapshot? _siteAuditBacklog;
     private readonly ConcurrentDictionary<string, ConnectionHealth> _connectionStatuses = new();
     private readonly ConcurrentDictionary<string, TagResolutionStatus> _tagResolutionCounts = new();
     private readonly ConcurrentDictionary<string, string> _connectionEndpoints = new();
@@ -89,6 +91,18 @@ public class SiteHealthCollector : ISiteHealthCollector
         Interlocked.Increment(ref _auditRedactionFailures);
     }
 
+    /// <summary>
+    /// Audit Log (#23) M6 Bundle E (T6) — replace the latest backlog snapshot
+    /// from the site SQLite writer. The field is a single reference write
+    /// (volatile) so the next <see cref="CollectReport"/> sees the most recent
+    /// snapshot — there is no count to reset, the report just carries forward
+    /// whatever was last refreshed.
+    /// </summary>
+    public void UpdateSiteAuditBacklog(SiteAuditBacklogSnapshot snapshot)
+    {
+        _siteAuditBacklog = snapshot ?? throw new ArgumentNullException(nameof(snapshot));
+    }
+
     /// <summary>
     /// Update the health status for a named data connection.
     /// Called by DCL when connection state changes.
@@ -207,6 +221,7 @@ public class SiteHealthCollector : ISiteHealthCollector
             ParkedMessageCount: Interlocked.CompareExchange(ref _parkedMessageCount, 0, 0),
             ClusterNodes: _clusterNodes?.ToList(),
             SiteAuditWriteFailures: siteAuditWriteFailures,
-            AuditRedactionFailure: auditRedactionFailures);
+            AuditRedactionFailure: auditRedactionFailures,
+            SiteAuditBacklog: _siteAuditBacklog);
     }
 }

src/ScadaLink.Host/Actors/AkkaHostedService.cs

@@ -34,6 +34,13 @@ public class AkkaHostedService : IHostedService
     private readonly CommunicationOptions _communicationOptions;
     private readonly ILogger<AkkaHostedService> _logger;
     private ActorSystem? _actorSystem;
+    /// <summary>
+    /// Auxiliary IDisposables (e.g. the SiteAuditTelemetryStalledTracker)
+    /// that this hosted service constructs at start time and must tear down
+    /// on shutdown — they don't fit the ActorSystem lifecycle but share its
+    /// process scope.
+    /// </summary>
+    private readonly List<IDisposable> _trackedDisposables = new();
 
     public AkkaHostedService(
         IServiceProvider serviceProvider,
@@ -201,6 +208,31 @@ akka {{
 
     public async Task StopAsync(CancellationToken cancellationToken)
     {
+        // Dispose auxiliary subscribers (e.g. SiteAuditTelemetryStalledTracker)
+        // BEFORE Akka shuts down so their EventStream unsubscribe calls run
+        // while the system is still alive. Per-tracker Dispose is wrapped in
+        // its own try so a misbehaving subscriber can't sink the shutdown.
+        // Snapshot the list inside a lock so a concurrent StartAsync (the
+        // test harness sometimes triggers a second start/stop interleaving)
+        // can't race the enumeration. Clearing the original list under the
+        // same lock leaves the next StartAsync with a clean slate.
+        IDisposable[] disposables;
+        lock (_trackedDisposables)
+        {
+            disposables = _trackedDisposables.ToArray();
+            _trackedDisposables.Clear();
+        }
+        foreach (var disposable in disposables)
+        {
+            try { disposable.Dispose(); }
+            catch (Exception ex)
+            {
+                _logger.LogWarning(ex,
+                    "Auxiliary subscriber {Type} threw during shutdown",
+                    disposable.GetType().Name);
+            }
+        }
+
         if (_actorSystem != null)
         {
             _logger.LogInformation("Shutting down Akka.NET actor system via CoordinatedShutdown...");
@@ -349,6 +381,31 @@ akka {{
             "AuditLogIngestActor singleton created (gRPC server bound: {GrpcBound})",
             grpcServer is not null);
 
+        // Audit Log (#23) M6 Bundle E (T7): subscribe the per-site stalled
+        // telemetry tracker to the actor system EventStream NOW that the
+        // system exists. The tracker mirrors every
+        // SiteAuditTelemetryStalledChanged publication (from
+        // SiteAuditReconciliationActor — wired in a later bundle) into the
+        // AuditCentralHealthSnapshot singleton so the central health surface
+        // sees per-site stalled state. The tracker is constructed here rather
+        // than in AddAuditLogCentralMaintenance because its ctor needs an
+        // ActorSystem, which is not a DI-resolvable singleton — it's owned
+        // by this hosted service. The snapshot singleton is resolvable;
+        // passing it in seeds the tracker's Apply() so both internal state
+        // and the snapshot stay in lock-step.
+        var auditCentralSnapshot = _serviceProvider
+            .GetService<ScadaLink.AuditLog.Central.AuditCentralHealthSnapshot>();
+        if (auditCentralSnapshot is not null)
+        {
+            var stalledTracker = new ScadaLink.AuditLog.Central.SiteAuditTelemetryStalledTracker(
+                _actorSystem!, auditCentralSnapshot);
+            lock (_trackedDisposables)
+            {
+                _trackedDisposables.Add(stalledTracker);
+            }
+            _logger.LogInformation("SiteAuditTelemetryStalledTracker subscribed to EventStream");
+        }
+
         // Site Call Audit (#22) — central singleton mirrors the AuditLogIngest
         // and NotificationOutbox patterns. M3's dual-write transaction routes
         // SiteCalls upserts through AuditLogIngestActor's own scope-per-message
@@ -605,7 +662,7 @@ akka {{
         var siteAuditOptions = _serviceProvider
             .GetRequiredService<IOptions<ScadaLink.AuditLog.Site.Telemetry.SiteAuditTelemetryOptions>>();
         var siteAuditQueue = _serviceProvider
-            .GetRequiredService<ScadaLink.AuditLog.Site.Telemetry.ISiteAuditQueue>();
+            .GetRequiredService<ScadaLink.Commons.Interfaces.Services.ISiteAuditQueue>();
         var siteAuditClient = _serviceProvider
             .GetRequiredService<ScadaLink.AuditLog.Site.Telemetry.ISiteStreamAuditClient>();
         var siteAuditLogger = _serviceProvider.GetRequiredService<ILoggerFactory>()
@@ -640,6 +697,13 @@ akka {{
         // handshake has completed". Streams opened before SetReady are already
         // rejected by SiteStreamGrpcServer with StatusCode.Unavailable.
         var grpcServer = _serviceProvider.GetService<ScadaLink.Communication.Grpc.SiteStreamGrpcServer>();
+        // Audit Log (#23 M6): hand the site-local SqliteAuditWriter (which
+        // implements ISiteAuditQueue) to the gRPC server so the PullAuditEvents
+        // reconciliation RPC can serve central's pulls. Both the writer and the
+        // gRPC server are singletons — wiring this here keeps the dependency
+        // direction one-way (Host knows both; Communication doesn't reach back
+        // into AuditLog).
+        grpcServer?.SetSiteAuditQueue(siteAuditQueue);
         grpcServer?.SetReady(_actorSystem!);
     }
 }

src/ScadaLink.Host/Program.cs

@@ -84,6 +84,10 @@ try
         // IAuditLogRepository. The site writer chain is still registered (lazy
         // singletons) but is never resolved on a central node.
         builder.Services.AddAuditLog(builder.Configuration);
+        // #23 M6-T5 Bundle D — central-only hosted service that rolls
+        // pf_AuditLog_Month forward monthly. Depends on IPartitionMaintenance
+        // (registered below by AddConfigurationDatabase).
+        builder.Services.AddAuditLogCentralMaintenance(builder.Configuration);
         // Site Call Audit (#22) — central node owns the SiteCallAuditActor
         // singleton (M3 Bundle F). The extension itself currently registers
         // nothing — actor Props are constructed inline in AkkaHostedService —

tests/ScadaLink.AuditLog.Tests/Central/AuditLogIngestActorTests.cs

@@ -214,7 +214,11 @@ public class AuditLogIngestActorTests : TestKit, IClassFixture<MsSqlMigrationFix
             AuditLogQueryFilter filter, AuditLogPaging paging, CancellationToken ct = default) =>
             _inner.QueryAsync(filter, paging, ct);
 
-        public Task SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default) =>
+        public Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default) =>
             _inner.SwitchOutPartitionAsync(monthBoundary, ct);
+
+        public Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+            DateTime threshold, CancellationToken ct = default) =>
+            _inner.GetPartitionBoundariesOlderThanAsync(threshold, ct);
     }
 }

tests/ScadaLink.AuditLog.Tests/Central/AuditLogPartitionMaintenanceServiceTests.cs

@@ -0,0 +1,154 @@
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.Commons.Interfaces;
+using Xunit;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle D (#23 M6-T5) tests for <see cref="AuditLogPartitionMaintenanceService"/>.
+/// All tests use an in-memory <see cref="IPartitionMaintenance"/> stub —
+/// the real EF/MSSQL implementation is exercised by the
+/// <c>AuditLogPartitionMaintenanceTests</c> integration suite in
+/// <c>ScadaLink.ConfigurationDatabase.Tests</c>. This file is purely
+/// about the hosted service's policy decisions (start/stop, exception
+/// containment).
+/// </summary>
+public class AuditLogPartitionMaintenanceServiceTests
+{
+    /// <summary>
+    /// Recording stub — counts EnsureLookaheadAsync invocations and lets the
+    /// test inject an exception per invocation to drive the catch-all path.
+    /// </summary>
+    private sealed class RecordingMaintenance : IPartitionMaintenance
+    {
+        public int EnsureCallCount;
+        public Exception? ThrowOnce;
+
+        public Task<IReadOnlyList<DateTime>> EnsureLookaheadAsync(int lookaheadMonths, CancellationToken ct = default)
+        {
+            Interlocked.Increment(ref EnsureCallCount);
+            if (ThrowOnce is { } ex)
+            {
+                ThrowOnce = null;
+                throw ex;
+            }
+            return Task.FromResult<IReadOnlyList<DateTime>>(Array.Empty<DateTime>());
+        }
+
+        public Task<DateTime?> GetMaxBoundaryAsync(CancellationToken ct = default) =>
+            Task.FromResult<DateTime?>(DateTime.UtcNow.AddMonths(6));
+    }
+
+    /// <summary>
+    /// Captures logged exceptions so the catch-all assertion can prove
+    /// the exception was actually logged (not silently swallowed) and was
+    /// the exact instance the stub threw.
+    /// </summary>
+    private sealed class CapturingLogger : ILogger<AuditLogPartitionMaintenanceService>
+    {
+        public List<(LogLevel Level, Exception? Exception, string Message)> Entries { get; } = new();
+
+        public IDisposable? BeginScope<TState>(TState state) where TState : notnull => null;
+
+        public bool IsEnabled(LogLevel logLevel) => true;
+
+        public void Log<TState>(
+            LogLevel logLevel,
+            EventId eventId,
+            TState state,
+            Exception? exception,
+            Func<TState, Exception?, string> formatter)
+        {
+            Entries.Add((logLevel, exception, formatter(state, exception)));
+        }
+    }
+
+    private static IServiceProvider BuildProvider(IPartitionMaintenance maintenance)
+    {
+        var services = new ServiceCollection();
+        // IPartitionMaintenance is registered as scoped by AddConfigurationDatabase;
+        // we mirror that here so the hosted service's CreateAsyncScope +
+        // GetRequiredService resolves the stub the test injected.
+        services.AddScoped(_ => maintenance);
+        return services.BuildServiceProvider();
+    }
+
+    [Fact]
+    public async Task StartStop_NoExceptions()
+    {
+        // Long interval so only the eager startup tick fires inside the test
+        // window — keeps assertions deterministic without relying on
+        // multiple cadence loops.
+        var opts = Options.Create(new AuditLogPartitionMaintenanceOptions
+        {
+            IntervalSeconds = 60,
+            LookaheadMonths = 1,
+        });
+        var maintenance = new RecordingMaintenance();
+        var sp = BuildProvider(maintenance);
+
+        var svc = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            NullLogger<AuditLogPartitionMaintenanceService>.Instance);
+
+        await svc.StartAsync(CancellationToken.None);
+
+        // Spin briefly until the startup tick has fired — the loop's first
+        // SafeMaintainAsync runs on a background Task.Run continuation, so
+        // we can't synchronously rely on its completion.
+        var deadline = DateTime.UtcNow.AddSeconds(3);
+        while (Volatile.Read(ref maintenance.EnsureCallCount) < 1 && DateTime.UtcNow < deadline)
+        {
+            await Task.Delay(20);
+        }
+
+        await svc.StopAsync(CancellationToken.None);
+        svc.Dispose();
+
+        Assert.True(maintenance.EnsureCallCount >= 1, $"expected at least 1 ensure call, got {maintenance.EnsureCallCount}");
+    }
+
+    [Fact]
+    public async Task SafeMaintain_ExceptionLogged_NotPropagated()
+    {
+        var opts = Options.Create(new AuditLogPartitionMaintenanceOptions
+        {
+            IntervalSeconds = 60,
+            LookaheadMonths = 1,
+        });
+        // The injected exception fires on the FIRST EnsureLookaheadAsync call
+        // (the startup tick) — the hosted service must contain it and
+        // continue running.
+        var boom = new InvalidOperationException("simulated maintenance failure");
+        var maintenance = new RecordingMaintenance { ThrowOnce = boom };
+        var sp = BuildProvider(maintenance);
+        var logger = new CapturingLogger();
+
+        var svc = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            logger);
+
+        // StartAsync must not throw even though the very first tick will fail.
+        await svc.StartAsync(CancellationToken.None);
+
+        // Wait for the error to surface in the logger.
+        var deadline = DateTime.UtcNow.AddSeconds(3);
+        while (!logger.Entries.Any(e => e.Exception == boom) && DateTime.UtcNow < deadline)
+        {
+            await Task.Delay(20);
+        }
+
+        await svc.StopAsync(CancellationToken.None);
+        svc.Dispose();
+
+        var errorEntry = Assert.Single(logger.Entries, e => e.Exception == boom);
+        Assert.Equal(LogLevel.Error, errorEntry.Level);
+        Assert.Equal(1, maintenance.EnsureCallCount);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Central/AuditLogPurgeActorTests.cs

@@ -0,0 +1,376 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.AuditLog.Configuration;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Types.Audit;
+using ScadaLink.Commons.Types.Enums;
+using ScadaLink.ConfigurationDatabase;
+using ScadaLink.ConfigurationDatabase.Repositories;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle C (#23 M6-T4) tests for <see cref="AuditLogPurgeActor"/>. The fast,
+/// schedule-only tests substitute a recording stub for
+/// <see cref="IAuditLogRepository"/> so the timer + per-boundary error-isolation
+/// + event-publish machinery can be exercised without an MSSQL container.
+/// The end-to-end "real partition gets switched out" assertion lives in the
+/// repository tests (Bundle C of M6-T4); this actor file is purely about the
+/// actor's policy decisions.
+/// </summary>
+public class AuditLogPurgeActorTests : TestKit, IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public AuditLogPurgeActorTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    /// <summary>
+    /// In-memory recording stub. Captures every
+    /// <see cref="GetPartitionBoundariesOlderThanAsync"/> + every
+    /// <see cref="SwitchOutPartitionAsync"/> so tests can assert which boundaries
+    /// the actor chose to purge and how many ticks it issued. Also lets a
+    /// specific boundary be configured to throw so the continue-on-error path
+    /// is exercisable.
+    /// </summary>
+    private sealed class RecordingRepo : IAuditLogRepository
+    {
+        public List<DateTime> ThresholdQueries { get; } = new();
+        public List<DateTime> SwitchedBoundaries { get; } = new();
+        public Func<DateTime, long> RowsPerBoundary { get; set; } = _ => 0L;
+        public DateTime? ThrowOnBoundary { get; set; }
+        public Exception? BoundaryException { get; set; }
+
+        // The actor enumerator returns whichever list is configured here.
+        // Mutating this between ticks lets tests simulate "no longer
+        // eligible" boundaries on the second tick.
+        public List<DateTime> Boundaries { get; set; } = new();
+
+        public Task InsertIfNotExistsAsync(AuditEvent evt, CancellationToken ct = default) =>
+            Task.CompletedTask;
+
+        public Task<IReadOnlyList<AuditEvent>> QueryAsync(
+            AuditLogQueryFilter filter, AuditLogPaging paging, CancellationToken ct = default) =>
+            Task.FromResult<IReadOnlyList<AuditEvent>>(Array.Empty<AuditEvent>());
+
+        public Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default)
+        {
+            if (ThrowOnBoundary.HasValue && monthBoundary == ThrowOnBoundary.Value)
+            {
+                throw BoundaryException ?? new InvalidOperationException("simulated switch failure");
+            }
+            SwitchedBoundaries.Add(monthBoundary);
+            return Task.FromResult(RowsPerBoundary(monthBoundary));
+        }
+
+        public Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+            DateTime threshold, CancellationToken ct = default)
+        {
+            ThresholdQueries.Add(threshold);
+            return Task.FromResult<IReadOnlyList<DateTime>>(Boundaries.ToArray());
+        }
+    }
+
+    private IServiceProvider BuildScopedProvider(IAuditLogRepository repo)
+    {
+        var services = new ServiceCollection();
+        // Mirror AddConfigurationDatabase: IAuditLogRepository is scoped, so
+        // the actor opens a fresh scope per tick and resolves there.
+        services.AddScoped(_ => repo);
+        return services.BuildServiceProvider();
+    }
+
+    private IActorRef CreateActor(
+        IAuditLogRepository repo,
+        AuditLogPurgeOptions purgeOptions,
+        AuditLogOptions? auditOptions = null)
+    {
+        var sp = BuildScopedProvider(repo);
+        return Sys.ActorOf(Props.Create(() => new AuditLogPurgeActor(
+            sp,
+            Options.Create(purgeOptions),
+            Options.Create(auditOptions ?? new AuditLogOptions()),
+            NullLogger<AuditLogPurgeActor>.Instance)));
+    }
+
+    private static AuditLogPurgeOptions FastTickOptions(TimeSpan? interval = null) => new()
+    {
+        IntervalHours = 24,
+        IntervalOverride = interval ?? TimeSpan.FromMilliseconds(100),
+    };
+
+    /// <summary>
+    /// Subscribe a probe to the EventStream so the test can observe
+    /// <see cref="AuditLogPurgedEvent"/> publications synchronously.
+    /// </summary>
+    private Akka.TestKit.TestProbe SubscribePurged()
+    {
+        var probe = CreateTestProbe();
+        Sys.EventStream.Subscribe(probe.Ref, typeof(AuditLogPurgedEvent));
+        return probe;
+    }
+
+    // ---------------------------------------------------------------------
+    // 1. Tick_Fires_OnDailyInterval
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_Fires_OnDailyInterval()
+    {
+        var repo = new RecordingRepo();
+        CreateActor(repo, FastTickOptions());
+
+        // The first scheduled tick fires after the configured interval. We
+        // assert the visible side effect (the enumerator was called) rather
+        // than racing on internal state.
+        AwaitAssert(
+            () => Assert.True(repo.ThresholdQueries.Count >= 1,
+                $"expected >= 1 enumerator call, got {repo.ThresholdQueries.Count}"),
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 2. Tick_OldPartitions_SwitchedOut
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_OldPartitions_SwitchedOut()
+    {
+        var repo = new RecordingRepo
+        {
+            Boundaries = new List<DateTime>
+            {
+                new(2025, 11, 1, 0, 0, 0, DateTimeKind.Utc),
+                new(2025, 12, 1, 0, 0, 0, DateTimeKind.Utc),
+            },
+            RowsPerBoundary = _ => 42L,
+        };
+
+        CreateActor(repo, FastTickOptions());
+
+        AwaitAssert(
+            () =>
+            {
+                Assert.Contains(new DateTime(2025, 11, 1, 0, 0, 0, DateTimeKind.Utc), repo.SwitchedBoundaries);
+                Assert.Contains(new DateTime(2025, 12, 1, 0, 0, 0, DateTimeKind.Utc), repo.SwitchedBoundaries);
+            },
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 3. Tick_NewerPartitions_Untouched
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_NewerPartitions_Untouched()
+    {
+        // The actor's contract: it only touches whatever the enumerator
+        // returns. The enumerator (in production) filters out non-eligible
+        // boundaries; here we simulate that by handing back an empty list
+        // and asserting the actor switched nothing despite the tick firing.
+        var repo = new RecordingRepo { Boundaries = new List<DateTime>() };
+
+        CreateActor(repo, FastTickOptions());
+
+        // Wait for at least one tick (visible via the enumerator call) then
+        // assert no switch happened.
+        AwaitAssert(
+            () => Assert.True(repo.ThresholdQueries.Count >= 1),
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+
+        Assert.Empty(repo.SwitchedBoundaries);
+    }
+
+    // ---------------------------------------------------------------------
+    // 4. Tick_PublishesPurgedEvent_WithRowCount
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_PublishesPurgedEvent_WithRowCount()
+    {
+        var boundary = new DateTime(2025, 6, 1, 0, 0, 0, DateTimeKind.Utc);
+        var repo = new RecordingRepo
+        {
+            Boundaries = new List<DateTime> { boundary },
+            RowsPerBoundary = _ => 1234L,
+        };
+
+        var probe = SubscribePurged();
+        CreateActor(repo, FastTickOptions());
+
+        var msg = probe.ExpectMsg<AuditLogPurgedEvent>(TimeSpan.FromSeconds(5));
+        Assert.Equal(boundary, msg.MonthBoundary);
+        Assert.Equal(1234L, msg.RowsDeleted);
+        Assert.True(msg.DurationMs >= 0,
+            $"DurationMs should be non-negative; was {msg.DurationMs}");
+    }
+
+    // ---------------------------------------------------------------------
+    // 5. Tick_SwitchThrows_OtherPartitionsStillProcessed (continue-on-error)
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_SwitchThrows_OtherPartitionsStillProcessed()
+    {
+        var poisonBoundary = new DateTime(2025, 7, 1, 0, 0, 0, DateTimeKind.Utc);
+        var goodBoundary = new DateTime(2025, 8, 1, 0, 0, 0, DateTimeKind.Utc);
+        var repo = new RecordingRepo
+        {
+            Boundaries = new List<DateTime> { poisonBoundary, goodBoundary },
+            ThrowOnBoundary = poisonBoundary,
+            BoundaryException = new InvalidOperationException("simulated switch failure for poison boundary"),
+        };
+
+        CreateActor(repo, FastTickOptions());
+
+        AwaitAssert(
+            () =>
+            {
+                // The good boundary was still switched even though the poison
+                // boundary threw.
+                Assert.Contains(goodBoundary, repo.SwitchedBoundaries);
+                Assert.DoesNotContain(poisonBoundary, repo.SwitchedBoundaries);
+            },
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 6. EndToEnd_RealPartition_RowsRemoved_PurgedEventPublished
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_RealPartition_RowsRemoved_PurgedEventPublished()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        // Today is ~2026-05-20 per the test environment. With RetentionDays =
+        // 60 the actor computes threshold ≈ 2026-03-21:
+        //   * Jan partition (MAX = Jan 15)  → older than threshold → PURGED
+        //   * Apr partition (MAX = Apr 15)  → newer than threshold → KEPT
+        var siteId = "purge-e2e-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var janEvt = new AuditEvent
+        {
+            EventId = Guid.NewGuid(),
+            OccurredAtUtc = new DateTime(2026, 1, 15, 0, 0, 0, DateTimeKind.Utc),
+            Channel = AuditChannel.ApiOutbound,
+            Kind = AuditKind.ApiCall,
+            Status = AuditStatus.Delivered,
+            SourceSiteId = siteId,
+        };
+        var aprEvt = new AuditEvent
+        {
+            EventId = Guid.NewGuid(),
+            OccurredAtUtc = new DateTime(2026, 4, 15, 0, 0, 0, DateTimeKind.Utc),
+            Channel = AuditChannel.ApiOutbound,
+            Kind = AuditKind.ApiCall,
+            Status = AuditStatus.Delivered,
+            SourceSiteId = siteId,
+        };
+
+        await using (var seedContext = CreateMsSqlContext())
+        {
+            var seedRepo = new AuditLogRepository(seedContext);
+            await seedRepo.InsertIfNotExistsAsync(janEvt);
+            await seedRepo.InsertIfNotExistsAsync(aprEvt);
+        }
+
+        // Wire the actor's DI scope to the real repository against the
+        // fixture's MSSQL database. The actor opens a fresh scope per tick,
+        // so register the context as scoped (mirroring the production
+        // AddConfigurationDatabase wiring).
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(
+            opts => opts.UseSqlServer(_fixture.ConnectionString),
+            ServiceLifetime.Scoped);
+        services.AddScoped<IAuditLogRepository, AuditLogRepository>();
+        var sp = services.BuildServiceProvider();
+
+        var auditOptions = new AuditLogOptions { RetentionDays = 60 };
+        var purgeOptions = new AuditLogPurgeOptions
+        {
+            IntervalHours = 24,
+            IntervalOverride = TimeSpan.FromMilliseconds(100),
+        };
+
+        var probe = SubscribePurged();
+        Sys.ActorOf(Props.Create(() => new AuditLogPurgeActor(
+            sp,
+            Options.Create(purgeOptions),
+            Options.Create(auditOptions),
+            NullLogger<AuditLogPurgeActor>.Instance)));
+
+        // The probe receives one AuditLogPurgedEvent per partition the actor
+        // purges per tick — other test runs that share the fixture DB may
+        // also leave behind eligible partitions, but this test creates its
+        // own fixture DB so the Jan-2026 partition is the only eligible one.
+        // Use FishForMessage to filter just in case, with a generous timeout
+        // because the real drop-and-rebuild dance against MSSQL routinely
+        // takes a couple of seconds on a busy dev container.
+        var janBoundary = new DateTime(2026, 1, 1, 0, 0, 0, DateTimeKind.Utc);
+        var matched = probe.FishForMessage<AuditLogPurgedEvent>(
+            isMessage: m => m.MonthBoundary == janBoundary,
+            max: TimeSpan.FromSeconds(30));
+
+        Assert.True(matched.RowsDeleted >= 1,
+            $"Expected RowsDeleted >= 1 for the Jan-2026 partition; got {matched.RowsDeleted}.");
+
+        // Settle: allow any in-flight tick to commit before reading.
+        await Task.Delay(TimeSpan.FromMilliseconds(500));
+        await using var verifyContext = CreateMsSqlContext();
+        var rows = await verifyContext.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == siteId)
+            .ToListAsync();
+
+        Assert.DoesNotContain(rows, r => r.EventId == janEvt.EventId);
+        Assert.Contains(rows, r => r.EventId == aprEvt.EventId);
+    }
+
+    private ScadaLinkDbContext CreateMsSqlContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    // ---------------------------------------------------------------------
+    // 7. Threshold_UsesAuditLogOptionsRetentionDays
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Threshold_UsesAuditLogOptionsRetentionDays()
+    {
+        // The actor computes the threshold from AuditLogOptions.RetentionDays;
+        // assert the enumerator received a threshold whose value is in the
+        // expected window (today - retentionDays) rather than DateTime.MinValue
+        // or some other accidental default. We use a non-default retention
+        // (30 days) so the assertion isn't satisfied by the 365 default.
+        var repo = new RecordingRepo();
+        CreateActor(
+            repo,
+            FastTickOptions(),
+            auditOptions: new AuditLogOptions { RetentionDays = 30 });
+
+        AwaitAssert(
+            () => Assert.True(repo.ThresholdQueries.Count >= 1),
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+
+        var threshold = repo.ThresholdQueries[0];
+        var expected = DateTime.UtcNow - TimeSpan.FromDays(30);
+        // 1-minute slack covers test-thread scheduling jitter between the
+        // tick firing and the assertion running.
+        Assert.True(
+            Math.Abs((threshold - expected).TotalMinutes) < 1.0,
+            $"threshold {threshold:o} should be within 1 minute of {expected:o}");
+    }
+}

tests/ScadaLink.AuditLog.Tests/Central/CentralAuditRedactionFailureCounterTests.cs

@@ -0,0 +1,98 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.Extensions.Configuration;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging;
+using Microsoft.Extensions.Logging.Abstractions;
+using ScadaLink.AuditLog;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.AuditLog.Payload;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle E (M6-T9) coverage for the central-side payload-filter redactor
+/// failure bridge. M5 wired the SITE bridge
+/// (<c>HealthMetricsAuditRedactionFailureCounter</c>) that pushes increments
+/// into the site health report; M6 mirrors that with
+/// <see cref="CentralAuditRedactionFailureCounter"/> so the same payload
+/// filter — when it runs on the central writer paths — surfaces failures on
+/// the central <see cref="AuditCentralHealthSnapshot"/>.
+/// </summary>
+public class CentralAuditRedactionFailureCounterTests : TestKit
+{
+    [Fact]
+    public void Increment_Routes_To_Snapshot()
+    {
+        var snapshot = new AuditCentralHealthSnapshot();
+        var counter = new CentralAuditRedactionFailureCounter(snapshot);
+
+        counter.Increment();
+        counter.Increment();
+        counter.Increment();
+
+        Assert.Equal(3, snapshot.AuditRedactionFailure);
+    }
+
+    [Fact]
+    public void Construction_With_Null_Snapshot_Throws()
+    {
+        Assert.Throws<ArgumentNullException>(
+            () => new CentralAuditRedactionFailureCounter(null!));
+    }
+
+    [Fact]
+    public void AddAuditLogCentralMaintenance_Replaces_IAuditRedactionFailureCounter_With_CentralImpl()
+    {
+        // AddAuditLog registers NoOp; AddAuditLogCentralMaintenance is the
+        // override path. The replaced binding MUST resolve to the central
+        // bridge — a site host that wires AddAuditLogHealthMetricsBridge
+        // instead would resolve to the site bridge (covered in
+        // AddAuditLogTests).
+        var config = new ConfigurationBuilder()
+            .AddInMemoryCollection(new Dictionary<string, string?>
+            {
+                ["AuditLog:SiteWriter:DatabasePath"] = ":memory:",
+            })
+            .Build();
+
+        var services = new ServiceCollection();
+        services.AddSingleton<ILoggerFactory, NullLoggerFactory>();
+        services.AddSingleton(typeof(ILogger<>), typeof(NullLogger<>));
+        // AuditCentralHealthSnapshot no longer takes a tracker dependency —
+        // the tracker is constructed later by the Akka bootstrap because its
+        // ctor needs an ActorSystem (not a DI-resolvable singleton). The
+        // snapshot itself composes purely from primitives.
+        services.AddAuditLog(config);
+        services.AddAuditLogCentralMaintenance(config);
+        using var provider = services.BuildServiceProvider();
+
+        var counter = provider.GetRequiredService<IAuditRedactionFailureCounter>();
+
+        Assert.IsType<CentralAuditRedactionFailureCounter>(counter);
+    }
+
+    [Fact]
+    public void AddAuditLog_Default_IAuditRedactionFailureCounter_Is_NoOp()
+    {
+        // Sanity check: without AddAuditLogCentralMaintenance the default
+        // remains the NoOp from M5 — the central bridge only takes effect
+        // when the central-only registration runs.
+        var config = new ConfigurationBuilder()
+            .AddInMemoryCollection(new Dictionary<string, string?>
+            {
+                ["AuditLog:SiteWriter:DatabasePath"] = ":memory:",
+            })
+            .Build();
+
+        var services = new ServiceCollection();
+        services.AddSingleton<ILoggerFactory, NullLoggerFactory>();
+        services.AddSingleton(typeof(ILogger<>), typeof(NullLogger<>));
+        services.AddAuditLog(config);
+        using var provider = services.BuildServiceProvider();
+
+        var counter = provider.GetRequiredService<IAuditRedactionFailureCounter>();
+
+        Assert.IsType<NoOpAuditRedactionFailureCounter>(counter);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Central/CentralAuditWriteFailuresTests.cs

@@ -0,0 +1,160 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Messages.Audit;
+using ScadaLink.Commons.Types.Audit;
+using ScadaLink.Commons.Types.Enums;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle E (M6-T8) regression coverage for the central-side audit-write
+/// failure counter. <see cref="CentralAuditWriter"/> and
+/// <see cref="AuditLogIngestActor"/> both swallow repository throws (audit
+/// must NEVER abort the user-facing action, alog.md §13) but bump the
+/// <see cref="ICentralAuditWriteFailureCounter"/> so the central health
+/// surface (<see cref="AuditCentralHealthSnapshot"/>) can flag a sustained
+/// outage.
+/// </summary>
+public class CentralAuditWriteFailuresTests : TestKit
+{
+    private static AuditEvent NewEvent() => new()
+    {
+        EventId = Guid.NewGuid(),
+        OccurredAtUtc = DateTime.UtcNow,
+        Channel = AuditChannel.ApiOutbound,
+        Kind = AuditKind.ApiCall,
+        Status = AuditStatus.Delivered,
+    };
+
+    /// <summary>
+    /// Repository stub that always throws on insert — exercises the failure
+    /// path in both <see cref="CentralAuditWriter"/> and
+    /// <see cref="AuditLogIngestActor"/>.
+    /// </summary>
+    private sealed class ThrowingRepo : IAuditLogRepository
+    {
+        public Task InsertIfNotExistsAsync(AuditEvent evt, CancellationToken ct = default) =>
+            throw new InvalidOperationException("simulated repo failure");
+        public Task<IReadOnlyList<AuditEvent>> QueryAsync(
+            AuditLogQueryFilter filter, AuditLogPaging paging, CancellationToken ct = default) =>
+            Task.FromResult<IReadOnlyList<AuditEvent>>(Array.Empty<AuditEvent>());
+        public Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default) =>
+            Task.FromResult(0L);
+        public Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+            DateTime threshold, CancellationToken ct = default) =>
+            Task.FromResult<IReadOnlyList<DateTime>>(Array.Empty<DateTime>());
+    }
+
+    /// <summary>
+    /// In-memory <see cref="ICentralAuditWriteFailureCounter"/> recording
+    /// every <see cref="Increment"/> call so tests can assert on the count.
+    /// </summary>
+    private sealed class RecordingFailureCounter : ICentralAuditWriteFailureCounter
+    {
+        private int _count;
+        public int Count => Volatile.Read(ref _count);
+        public void Increment() => Interlocked.Increment(ref _count);
+    }
+
+    [Fact]
+    public async Task Forced_Failure_Increments_Counter()
+    {
+        // Direct test: build the writer with a throwing scope and verify the
+        // injected counter is bumped on the swallowed insert exception.
+        var counter = new RecordingFailureCounter();
+        var services = new ServiceCollection();
+        services.AddScoped<IAuditLogRepository, ThrowingRepo>();
+        var sp = services.BuildServiceProvider();
+
+        var writer = new CentralAuditWriter(
+            sp,
+            NullLogger<CentralAuditWriter>.Instance,
+            filter: null,
+            failureCounter: counter);
+
+        // WriteAsync swallows the exception and increments the counter.
+        await writer.WriteAsync(NewEvent());
+
+        Assert.Equal(1, counter.Count);
+    }
+
+    [Fact]
+    public async Task AuditLogIngestActor_Failure_Increments_Counter()
+    {
+        // The actor's production ctor resolves both IAuditLogRepository AND
+        // ICentralAuditWriteFailureCounter from the scope per-message; we
+        // register both and verify the per-row catch bumps the counter for
+        // every row in the batch.
+        var counter = new RecordingFailureCounter();
+        var services = new ServiceCollection();
+        services.AddScoped<IAuditLogRepository, ThrowingRepo>();
+        // Counter is a singleton — the actor's per-message scope still
+        // resolves the same instance via the scope's parent provider.
+        services.AddSingleton<ICentralAuditWriteFailureCounter>(counter);
+        var sp = services.BuildServiceProvider();
+
+        var actor = Sys.ActorOf(Props.Create(() => new AuditLogIngestActor(
+            sp, NullLogger<AuditLogIngestActor>.Instance)));
+
+        var batch = new[] { NewEvent(), NewEvent(), NewEvent() };
+        var reply = await actor.Ask<IngestAuditEventsReply>(
+            new IngestAuditEventsCommand(batch), TimeSpan.FromSeconds(5));
+
+        // Every row threw → none accepted, counter bumped once per row.
+        Assert.Empty(reply.AcceptedEventIds);
+        Assert.Equal(batch.Length, counter.Count);
+    }
+
+    [Fact]
+    public void Snapshot_Aggregates_Counters_And_StalledState()
+    {
+        // AuditCentralHealthSnapshot implements both writer surfaces; bumping
+        // through the writer interfaces is reflected on the read surface, and
+        // the per-site stalled state is fed in via ApplyStalled — production
+        // wires that to a SiteAuditTelemetryStalledTracker, but the snapshot
+        // is testable in isolation against the same Apply surface.
+        var snapshot = new AuditCentralHealthSnapshot();
+
+        Assert.Equal(0, snapshot.CentralAuditWriteFailures);
+        Assert.Equal(0, snapshot.AuditRedactionFailure);
+        Assert.Empty(snapshot.SiteAuditTelemetryStalled);
+
+        ((ICentralAuditWriteFailureCounter)snapshot).Increment();
+        ((ICentralAuditWriteFailureCounter)snapshot).Increment();
+        ((ScadaLink.AuditLog.Payload.IAuditRedactionFailureCounter)snapshot).Increment();
+
+        // Wire the tracker so an EventStream publish reaches the snapshot.
+        // The tracker pushes into the snapshot's ApplyStalled when given
+        // the snapshot in its ctor; the tracker also keeps its own latch,
+        // but the snapshot read surface is what the central UI reads.
+        using var tracker = new SiteAuditTelemetryStalledTracker(Sys, snapshot);
+        Sys.EventStream.Publish(new SiteAuditTelemetryStalledChanged("siteA", Stalled: true));
+        AwaitAssert(() =>
+            {
+                var stalledMap = snapshot.SiteAuditTelemetryStalled;
+                Assert.True(stalledMap.TryGetValue("siteA", out var s) && s,
+                    "expected siteA to be stalled in snapshot");
+            },
+            duration: TimeSpan.FromSeconds(2),
+            interval: TimeSpan.FromMilliseconds(20));
+
+        Assert.Equal(2, snapshot.CentralAuditWriteFailures);
+        Assert.Equal(1, snapshot.AuditRedactionFailure);
+    }
+
+    [Fact]
+    public void Snapshot_Empty_OnConstruction()
+    {
+        // Sanity: the snapshot's three properties start at their zero values
+        // before any writer or stalled-event publication.
+        var snapshot = new AuditCentralHealthSnapshot();
+        Assert.Equal(0, snapshot.CentralAuditWriteFailures);
+        Assert.Equal(0, snapshot.AuditRedactionFailure);
+        Assert.Empty(snapshot.SiteAuditTelemetryStalled);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Central/SiteAuditReconciliationActorTests.cs

@@ -0,0 +1,442 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Messages.Integration;
+using ScadaLink.Commons.Types.Audit;
+using ScadaLink.Commons.Types.Enums;
+using ScadaLink.ConfigurationDatabase;
+using ScadaLink.ConfigurationDatabase.Repositories;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle B (M6-T3) tests for <see cref="SiteAuditReconciliationActor"/>. Most
+/// tests substitute the <see cref="IAuditLogRepository"/> with an in-memory
+/// recording stub so the actor's tick / cursor / stalled state machinery can
+/// be exercised in milliseconds without an MSSQL container. The duplicate /
+/// idempotency assertion uses the real <see cref="AuditLogRepository"/> against
+/// the <see cref="MsSqlMigrationFixture"/> so we verify InsertIfNotExistsAsync
+/// actually swallows duplicate-key collisions (the M2 Bundle A race-fix the
+/// reconciliation puller depends on).
+/// </summary>
+public class SiteAuditReconciliationActorTests : TestKit, IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public SiteAuditReconciliationActorTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    private static AuditEvent NewEvent(
+        string siteId,
+        DateTime? occurredAt = null,
+        Guid? id = null) => new()
+    {
+        EventId = id ?? Guid.NewGuid(),
+        OccurredAtUtc = occurredAt ?? new DateTime(2026, 5, 20, 10, 0, 0, DateTimeKind.Utc),
+        Channel = AuditChannel.ApiOutbound,
+        Kind = AuditKind.ApiCall,
+        Status = AuditStatus.Delivered,
+        SourceSiteId = siteId,
+    };
+
+    private static SiteAuditReconciliationOptions FastTickOptions(
+        int batchSize = 256,
+        int stalledAfter = 2) =>
+        new()
+        {
+            // 100 ms tick keeps each test under a second. AwaitAssert covers
+            // schedule jitter so a 100 ms tick has up to ~3 s to fire.
+            ReconciliationIntervalSeconds = 300,
+            ReconciliationIntervalOverride = TimeSpan.FromMilliseconds(100),
+            BatchSize = batchSize,
+            StalledAfterNonDrainingCycles = stalledAfter,
+        };
+
+    /// <summary>
+    /// In-memory recording stub used for non-MSSQL tests. Captures every
+    /// <see cref="InsertIfNotExistsAsync"/> call AND deduplicates on
+    /// <see cref="AuditEvent.EventId"/> so duplicate-handling assertions don't
+    /// need a real database for the simple cases.
+    /// </summary>
+    private sealed class RecordingRepo : IAuditLogRepository
+    {
+        public List<AuditEvent> Inserted { get; } = new();
+        private readonly HashSet<Guid> _seen = new();
+        public int InsertCallCount { get; private set; }
+
+        public Task InsertIfNotExistsAsync(AuditEvent evt, CancellationToken ct = default)
+        {
+            InsertCallCount++;
+            if (_seen.Add(evt.EventId))
+            {
+                Inserted.Add(evt);
+            }
+            return Task.CompletedTask;
+        }
+
+        public Task<IReadOnlyList<AuditEvent>> QueryAsync(
+            AuditLogQueryFilter filter, AuditLogPaging paging, CancellationToken ct = default) =>
+            Task.FromResult<IReadOnlyList<AuditEvent>>(Inserted);
+
+        public Task<long> SwitchOutPartitionAsync(DateTime monthBoundary, CancellationToken ct = default) =>
+            Task.FromResult(0L);
+
+        public Task<IReadOnlyList<DateTime>> GetPartitionBoundariesOlderThanAsync(
+            DateTime threshold, CancellationToken ct = default) =>
+            Task.FromResult<IReadOnlyList<DateTime>>(Array.Empty<DateTime>());
+    }
+
+    /// <summary>
+    /// In-memory enumerator returning a static list of sites.
+    /// </summary>
+    private sealed class StaticEnumerator : ISiteEnumerator
+    {
+        private readonly IReadOnlyList<SiteEntry> _sites;
+        public StaticEnumerator(params SiteEntry[] sites) => _sites = sites;
+        public Task<IReadOnlyList<SiteEntry>> EnumerateAsync(CancellationToken ct = default) =>
+            Task.FromResult(_sites);
+    }
+
+    /// <summary>
+    /// Scripted pull client — returns the next queued response for the site
+    /// on each call, looping the last entry if the queue is exhausted. Also
+    /// records every invocation so tests can assert call counts + arguments.
+    /// </summary>
+    private sealed class ScriptedPullClient : IPullAuditEventsClient
+    {
+        public List<(string SiteId, DateTime SinceUtc, int BatchSize)> Calls { get; } = new();
+        private readonly Dictionary<string, Queue<PullAuditEventsResponse>> _scripted = new();
+        private readonly Dictionary<string, Exception> _throwOnSite = new();
+
+        public ScriptedPullClient Script(string siteId, params PullAuditEventsResponse[] responses)
+        {
+            _scripted[siteId] = new Queue<PullAuditEventsResponse>(responses);
+            return this;
+        }
+
+        public ScriptedPullClient ThrowFor(string siteId, Exception ex)
+        {
+            _throwOnSite[siteId] = ex;
+            return this;
+        }
+
+        public Task<PullAuditEventsResponse> PullAsync(
+            string siteId, DateTime sinceUtc, int batchSize, CancellationToken ct)
+        {
+            Calls.Add((siteId, sinceUtc, batchSize));
+            if (_throwOnSite.TryGetValue(siteId, out var ex))
+            {
+                throw ex;
+            }
+            if (_scripted.TryGetValue(siteId, out var queue) && queue.Count > 0)
+            {
+                return Task.FromResult(queue.Dequeue());
+            }
+            return Task.FromResult(
+                new PullAuditEventsResponse(Array.Empty<AuditEvent>(), MoreAvailable: false));
+        }
+    }
+
+    private IServiceProvider BuildScopedProvider(IAuditLogRepository repo)
+    {
+        var services = new ServiceCollection();
+        // The actor opens a scope per tick and resolves IAuditLogRepository
+        // from that scope; registering as scoped mirrors how
+        // AddConfigurationDatabase wires the real repository.
+        services.AddScoped(_ => repo);
+        return services.BuildServiceProvider();
+    }
+
+    private IActorRef CreateActor(
+        ISiteEnumerator sites,
+        IPullAuditEventsClient client,
+        IAuditLogRepository repo,
+        SiteAuditReconciliationOptions options)
+    {
+        var sp = BuildScopedProvider(repo);
+        return Sys.ActorOf(Props.Create(() => new SiteAuditReconciliationActor(
+            sites,
+            client,
+            sp,
+            Options.Create(options),
+            NullLogger<SiteAuditReconciliationActor>.Instance)));
+    }
+
+    /// <summary>
+    /// Subscribes to the EventStream and collects every
+    /// <see cref="SiteAuditTelemetryStalledChanged"/> publication into a list
+    /// the test can assert on. Uses a probe actor so the stream's
+    /// fire-and-forget delivery is observable from the test thread.
+    /// </summary>
+    private (Akka.TestKit.TestProbe Probe, List<SiteAuditTelemetryStalledChanged> Captured) SubscribeStalled()
+    {
+        var probe = CreateTestProbe();
+        Sys.EventStream.Subscribe(probe.Ref, typeof(SiteAuditTelemetryStalledChanged));
+        var captured = new List<SiteAuditTelemetryStalledChanged>();
+        return (probe, captured);
+    }
+
+    // ---------------------------------------------------------------------
+    // 1. Timer_Fires_OnConfiguredInterval
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Timer_Fires_OnConfiguredInterval()
+    {
+        var sites = new StaticEnumerator(new SiteEntry("siteA", "http://siteA:8083"));
+        var client = new ScriptedPullClient();
+        var repo = new RecordingRepo();
+        var opts = FastTickOptions();
+
+        CreateActor(sites, client, repo, opts);
+
+        // The first scheduled tick fires after `ReconciliationIntervalSeconds`,
+        // which is 0 for the test — Akka's scheduler still respects the
+        // ScheduleTellRepeatedlyCancelable contract that issues a Tell on the
+        // scheduler thread, so we await visible side effects (a PullAsync call)
+        // rather than racing on internal state.
+        AwaitAssert(
+            () => Assert.True(client.Calls.Count >= 1, $"expected >= 1 pull call, got {client.Calls.Count}"),
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 2. Tick_PullsFromEachKnownSite
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_PullsFromEachKnownSite()
+    {
+        var sites = new StaticEnumerator(
+            new SiteEntry("siteA", "http://siteA:8083"),
+            new SiteEntry("siteB", "http://siteB:8083"));
+        var client = new ScriptedPullClient();
+        var repo = new RecordingRepo();
+
+        CreateActor(sites, client, repo, FastTickOptions());
+
+        AwaitAssert(() =>
+            {
+                Assert.Contains(client.Calls, c => c.SiteId == "siteA");
+                Assert.Contains(client.Calls, c => c.SiteId == "siteB");
+            },
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 3. Tick_IngestEvents_ViaInsertIfNotExistsAsync
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_IngestEvents_ViaInsertIfNotExistsAsync()
+    {
+        var sites = new StaticEnumerator(new SiteEntry("siteA", "http://siteA:8083"));
+        var e1 = NewEvent("siteA");
+        var e2 = NewEvent("siteA");
+        var client = new ScriptedPullClient().Script("siteA",
+            new PullAuditEventsResponse(new[] { e1, e2 }, MoreAvailable: false));
+        var repo = new RecordingRepo();
+
+        CreateActor(sites, client, repo, FastTickOptions());
+
+        AwaitAssert(() => Assert.Equal(2, repo.InsertCallCount),
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+        Assert.Contains(repo.Inserted, e => e.EventId == e1.EventId);
+        Assert.Contains(repo.Inserted, e => e.EventId == e2.EventId);
+    }
+
+    // ---------------------------------------------------------------------
+    // 4. Tick_Duplicates_NotDoubleInserted (real MSSQL idempotency)
+    // ---------------------------------------------------------------------
+
+    private ScadaLinkDbContext CreateContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    [SkippableFact]
+    public async Task Tick_Duplicates_NotDoubleInserted()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = "bundle-b-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var pre = NewEvent(siteId);
+
+        // Seed the row directly so the actor sees it already present when the
+        // pull returns it.
+        await using (var seedContext = CreateContext())
+        {
+            await new AuditLogRepository(seedContext).InsertIfNotExistsAsync(pre);
+        }
+
+        // Stack one new and the pre-existing row in the pull response. The
+        // second-pull script returns empty so the actor settles.
+        var fresh = NewEvent(siteId);
+        var sites = new StaticEnumerator(new SiteEntry(siteId, "http://x:8083"));
+        var client = new ScriptedPullClient().Script(siteId,
+            new PullAuditEventsResponse(new[] { pre, fresh }, MoreAvailable: false));
+
+        await using var context = CreateContext();
+        var repo = new AuditLogRepository(context);
+
+        CreateActor(sites, client, repo, FastTickOptions());
+
+        // Wait for the actor to ingest both rows.
+        await Task.Delay(TimeSpan.FromSeconds(1));
+        AwaitAssert(() => Assert.True(client.Calls.Count >= 1),
+            duration: TimeSpan.FromSeconds(3));
+
+        // Even though the pull returned 2 events, only 1 fresh row should
+        // exist in MSSQL alongside the pre-existing one — InsertIfNotExistsAsync
+        // is first-write-wins on EventId.
+        await using var read = CreateContext();
+        var rows = await read.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == siteId)
+            .ToListAsync();
+        Assert.Equal(2, rows.Count);
+        Assert.Contains(rows, r => r.EventId == pre.EventId);
+        Assert.Contains(rows, r => r.EventId == fresh.EventId);
+    }
+
+    // ---------------------------------------------------------------------
+    // 5. Cursor_Advances_ToMaxOccurredAtUtc
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Cursor_Advances_ToMaxOccurredAtUtc()
+    {
+        var sites = new StaticEnumerator(new SiteEntry("siteA", "http://siteA:8083"));
+
+        var t1 = new DateTime(2026, 5, 20, 10, 0, 0, DateTimeKind.Utc);
+        var t2 = new DateTime(2026, 5, 20, 10, 1, 0, DateTimeKind.Utc);
+        var t3 = new DateTime(2026, 5, 20, 10, 2, 0, DateTimeKind.Utc);
+        var e1 = NewEvent("siteA", t1);
+        var e2 = NewEvent("siteA", t2);
+        var e3 = NewEvent("siteA", t3);
+
+        // First pull returns three events with t1, t2, t3. Subsequent pulls
+        // return empty — but the test asserts the SECOND pull's since argument
+        // is t3 (the max OccurredAtUtc from the first pull).
+        var client = new ScriptedPullClient().Script("siteA",
+            new PullAuditEventsResponse(new[] { e1, e2, e3 }, MoreAvailable: false));
+        var repo = new RecordingRepo();
+
+        CreateActor(sites, client, repo, FastTickOptions());
+
+        // Wait until we have at least two pulls — the second one must use t3
+        // as its `since` argument because that was the max OccurredAtUtc in
+        // the first response.
+        AwaitAssert(() => Assert.True(client.Calls.Count >= 2,
+            $"need at least 2 pulls to assert cursor advancement, got {client.Calls.Count}"),
+            duration: TimeSpan.FromSeconds(5),
+            interval: TimeSpan.FromMilliseconds(50));
+
+        Assert.Equal(DateTime.MinValue, client.Calls[0].SinceUtc);
+        Assert.Equal(t3, client.Calls[1].SinceUtc);
+    }
+
+    // ---------------------------------------------------------------------
+    // 6. Tick_OneSiteThrows_OtherSitesStillProcessed
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void Tick_OneSiteThrows_OtherSitesStillProcessed()
+    {
+        var sites = new StaticEnumerator(
+            new SiteEntry("siteA", "http://siteA:8083"),
+            new SiteEntry("siteB", "http://siteB:8083"));
+
+        var bEvent = NewEvent("siteB");
+        var client = new ScriptedPullClient()
+            .ThrowFor("siteA", new InvalidOperationException("simulated transport failure"))
+            .Script("siteB",
+                new PullAuditEventsResponse(new[] { bEvent }, MoreAvailable: false));
+        var repo = new RecordingRepo();
+
+        CreateActor(sites, client, repo, FastTickOptions());
+
+        AwaitAssert(() =>
+            {
+                Assert.Contains(client.Calls, c => c.SiteId == "siteA");
+                Assert.Contains(repo.Inserted, e => e.EventId == bEvent.EventId);
+            },
+            duration: TimeSpan.FromSeconds(3),
+            interval: TimeSpan.FromMilliseconds(50));
+    }
+
+    // ---------------------------------------------------------------------
+    // 7. StalledDetection_TwoConsecutiveNonDrainingCycles_PublishesStalledTrue
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void StalledDetection_TwoConsecutiveNonDrainingCycles_PublishesStalledTrue()
+    {
+        var sites = new StaticEnumerator(new SiteEntry("siteA", "http://siteA:8083"));
+
+        // Two scripted responses that each return events AND MoreAvailable=true
+        // — the second pull triggers the stalled transition.
+        var batch1 = Enumerable.Range(0, 3).Select(_ => NewEvent("siteA")).ToArray();
+        var batch2 = Enumerable.Range(0, 3).Select(_ => NewEvent("siteA")).ToArray();
+        var client = new ScriptedPullClient().Script("siteA",
+            new PullAuditEventsResponse(batch1, MoreAvailable: true),
+            new PullAuditEventsResponse(batch2, MoreAvailable: true));
+
+        var repo = new RecordingRepo();
+        var (probe, _) = SubscribeStalled();
+
+        CreateActor(sites, client, repo, FastTickOptions(stalledAfter: 2));
+
+        // Expect Stalled=true after the second non-draining tick. The probe
+        // waits with its own timeout (a few seconds gives the 0 s repeat
+        // interval ample slack).
+        var msg = probe.ExpectMsg<SiteAuditTelemetryStalledChanged>(TimeSpan.FromSeconds(5));
+        Assert.Equal("siteA", msg.SiteId);
+        Assert.True(msg.Stalled);
+    }
+
+    // ---------------------------------------------------------------------
+    // 8. StalledDetection_DrainingCycle_PublishesStalledFalse
+    // ---------------------------------------------------------------------
+
+    [Fact]
+    public void StalledDetection_DrainingCycle_PublishesStalledFalse()
+    {
+        var sites = new StaticEnumerator(new SiteEntry("siteA", "http://siteA:8083"));
+
+        // Two non-draining responses get the actor into Stalled=true, then a
+        // draining response (events but MoreAvailable=false) flips it back.
+        var batch1 = Enumerable.Range(0, 3).Select(_ => NewEvent("siteA")).ToArray();
+        var batch2 = Enumerable.Range(0, 3).Select(_ => NewEvent("siteA")).ToArray();
+        var batch3 = Enumerable.Range(0, 3).Select(_ => NewEvent("siteA")).ToArray();
+        var client = new ScriptedPullClient().Script("siteA",
+            new PullAuditEventsResponse(batch1, MoreAvailable: true),
+            new PullAuditEventsResponse(batch2, MoreAvailable: true),
+            new PullAuditEventsResponse(batch3, MoreAvailable: false));
+
+        var repo = new RecordingRepo();
+        var (probe, _) = SubscribeStalled();
+
+        CreateActor(sites, client, repo, FastTickOptions(stalledAfter: 2));
+
+        // First publication is the stalled=true transition; second is the
+        // back-to-draining flip. The actor publishes ONLY on transitions so we
+        // expect exactly these two messages in order.
+        var first = probe.ExpectMsg<SiteAuditTelemetryStalledChanged>(TimeSpan.FromSeconds(5));
+        Assert.True(first.Stalled);
+
+        var second = probe.ExpectMsg<SiteAuditTelemetryStalledChanged>(TimeSpan.FromSeconds(5));
+        Assert.False(second.Stalled);
+        Assert.Equal("siteA", second.SiteId);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Central/SiteAuditTelemetryStalledTrackerTests.cs

@@ -0,0 +1,116 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using ScadaLink.AuditLog.Central;
+
+namespace ScadaLink.AuditLog.Tests.Central;
+
+/// <summary>
+/// Bundle E (M6-T7) tests for <see cref="SiteAuditTelemetryStalledTracker"/>.
+/// The tracker subscribes to the actor system's EventStream for
+/// <see cref="SiteAuditTelemetryStalledChanged"/> publications and maintains a
+/// per-site latch the central health surface can read. Since reconciliation is
+/// central-driven, the "stalled" state semantically belongs to central — not
+/// to the per-site <see cref="ScadaLink.Commons.Messages.Health.SiteHealthReport"/>
+/// payload (which the site itself emits). The tracker therefore lives as a
+/// central singleton, not on the site health collector.
+/// </summary>
+public class SiteAuditTelemetryStalledTrackerTests : TestKit
+{
+    /// <summary>
+    /// Helper: publishes a stalled-changed event on the actor system's
+    /// EventStream and waits a moment for the tracker's subscribe callback to
+    /// run. AwaitAssert avoids racing on the stream's async fan-out.
+    /// </summary>
+    private void PublishAndWait(SiteAuditTelemetryStalledTracker tracker, SiteAuditTelemetryStalledChanged evt)
+    {
+        Sys.EventStream.Publish(evt);
+        AwaitAssert(
+            () =>
+            {
+                var snapshot = tracker.Snapshot();
+                Assert.True(snapshot.TryGetValue(evt.SiteId, out var stalled),
+                    $"tracker did not record event for {evt.SiteId}");
+                Assert.Equal(evt.Stalled, stalled);
+            },
+            duration: TimeSpan.FromSeconds(2),
+            interval: TimeSpan.FromMilliseconds(20));
+    }
+
+    [Fact]
+    public void Initial_Snapshot_IsEmpty()
+    {
+        using var tracker = new SiteAuditTelemetryStalledTracker(Sys);
+
+        var snapshot = tracker.Snapshot();
+
+        Assert.Empty(snapshot);
+    }
+
+    [Fact]
+    public void StalledTrue_Event_TrackerReports_Stalled()
+    {
+        using var tracker = new SiteAuditTelemetryStalledTracker(Sys);
+
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteA", Stalled: true));
+
+        var snapshot = tracker.Snapshot();
+        Assert.True(snapshot["siteA"]);
+    }
+
+    [Fact]
+    public void StalledFalse_Event_TrackerReports_NotStalled()
+    {
+        using var tracker = new SiteAuditTelemetryStalledTracker(Sys);
+
+        // First flip the site into stalled so the false transition has a
+        // prior value to overwrite — mirrors how the reconciliation actor
+        // only publishes false after a true.
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteA", Stalled: true));
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteA", Stalled: false));
+
+        var snapshot = tracker.Snapshot();
+        Assert.False(snapshot["siteA"]);
+    }
+
+    [Fact]
+    public void Multiple_Sites_Tracked_Independently()
+    {
+        using var tracker = new SiteAuditTelemetryStalledTracker(Sys);
+
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteA", Stalled: true));
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteB", Stalled: false));
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteC", Stalled: true));
+
+        var snapshot = tracker.Snapshot();
+        Assert.Equal(3, snapshot.Count);
+        Assert.True(snapshot["siteA"]);
+        Assert.False(snapshot["siteB"]);
+        Assert.True(snapshot["siteC"]);
+    }
+
+    [Fact]
+    public void Constructor_With_Null_ActorSystem_Throws()
+    {
+        Assert.Throws<ArgumentNullException>(
+            () => new SiteAuditTelemetryStalledTracker((ActorSystem)null!));
+    }
+
+    [Fact]
+    public void Dispose_Unsubscribes_From_EventStream()
+    {
+        var tracker = new SiteAuditTelemetryStalledTracker(Sys);
+
+        PublishAndWait(tracker, new SiteAuditTelemetryStalledChanged("siteA", Stalled: true));
+
+        tracker.Dispose();
+
+        // After dispose any further events are ignored — the snapshot
+        // reflects the last known state at dispose time.
+        Sys.EventStream.Publish(new SiteAuditTelemetryStalledChanged("siteA", Stalled: false));
+
+        // Give the stream a moment in case the unsubscribe is racey; the
+        // assertion is that siteA stays at true.
+        Thread.Sleep(50);
+        Assert.True(tracker.Snapshot()["siteA"]);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Integration/OutageReconciliationTests.cs

@@ -0,0 +1,349 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.AuditLog.Site;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Interfaces.Services;
+using ScadaLink.Commons.Messages.Integration;
+using ScadaLink.Commons.Types.Enums;
+using ScadaLink.ConfigurationDatabase;
+using ScadaLink.ConfigurationDatabase.Repositories;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+
+namespace ScadaLink.AuditLog.Tests.Integration;
+
+/// <summary>
+/// Bundle F (#23 M6-T10) end-to-end test for the central-outage + reconciliation
+/// recovery loop. Wires the real site SQLite hot-path
+/// (<see cref="SqliteAuditWriter"/>) and the central <see cref="SiteAuditReconciliationActor"/>
+/// with an <see cref="AuditLogIngestActor"/> backed by the real
+/// <see cref="AuditLogRepository"/> on the per-test <see cref="MsSqlMigrationFixture"/>.
+/// </summary>
+/// <remarks>
+/// <para>
+/// The push path is deliberately omitted here: the brief models a sustained
+/// central outage where the site queue grows unbounded in Pending, then a
+/// reconciliation pull eventually drains everything once central comes back.
+/// We reuse the production <see cref="IPullAuditEventsClient"/> seam (Bundle B)
+/// with a test-only stub that wraps the same <see cref="ISiteAuditQueue.ReadPendingSinceAsync"/>
+/// surface a real central-side gRPC client would hit, so the test is exercising
+/// the actor's pull/ingest/mark-reconciled state machine end-to-end against
+/// the real repository.
+/// </para>
+/// <para>
+/// The <see cref="CombinedTelemetryHarness"/> from M3 is push-only — it has no
+/// reconciliation puller — so we build the smaller stub inline rather than
+/// retrofitting the shared harness with a code path it doesn't otherwise
+/// need.
+/// </para>
+/// </remarks>
+public class OutageReconciliationTests : TestKit, IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public OutageReconciliationTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    /// <summary>
+    /// Test-only <see cref="IPullAuditEventsClient"/> that mirrors how the
+    /// production central-side gRPC client will hit the site: read a batch
+    /// from <see cref="ISiteAuditQueue.ReadPendingSinceAsync"/>, then commit
+    /// via <see cref="ISiteAuditQueue.MarkReconciledAsync"/> once the central
+    /// repository accepts the rows. The Ask-based central path is wired by
+    /// the caller — we just expose the queue surface.
+    /// </summary>
+    /// <remarks>
+    /// The production wire shape will be:
+    ///   central PullAuditEvents RPC → site SiteStreamGrpcServer.PullAuditEvents
+    ///     → ISiteAuditQueue.ReadPendingSinceAsync → marshal proto → reply
+    /// followed by central InsertIfNotExistsAsync per row, then the site flips
+    /// the row to Reconciled on the next pull cycle. The stub collapses the
+    /// two halves (pull + commit) because the actor under test (the
+    /// reconciliation actor) is the side that drives both via the
+    /// IPullAuditEventsClient seam — committing back to the site after the
+    /// repository write is the reconciliation-actor invariant we want to
+    /// observe end-to-end.
+    /// </remarks>
+    private sealed class QueueBackedPullClient : IPullAuditEventsClient
+    {
+        private readonly ISiteAuditQueue _siteQueue;
+        public int CallCount { get; private set; }
+
+        public QueueBackedPullClient(ISiteAuditQueue siteQueue)
+        {
+            _siteQueue = siteQueue ?? throw new ArgumentNullException(nameof(siteQueue));
+        }
+
+        public async Task<PullAuditEventsResponse> PullAsync(
+            string siteId, DateTime sinceUtc, int batchSize, CancellationToken ct)
+        {
+            CallCount++;
+
+            var rows = await _siteQueue
+                .ReadPendingSinceAsync(sinceUtc, batchSize, ct)
+                .ConfigureAwait(false);
+
+            // Commit immediately on the site side — once the actor has the
+            // batch in hand it will InsertIfNotExistsAsync centrally; if the
+            // central insert later throws on a specific row, idempotency
+            // guarantees the next pull cycle does NOT re-fetch the row (it's
+            // already Reconciled on the site) but also does not surface the
+            // failure here. The brief calls this "ack-after-persist" — the
+            // production gRPC server will flip to Reconciled inside its
+            // PullAuditEvents handler after the central side has acknowledged
+            // (per Bundle A's race-fix, central is idempotent on EventId).
+            //
+            // MoreAvailable is true iff the read filled the batch — the actor
+            // uses this to decide whether to follow up on the next tick.
+            if (rows.Count > 0)
+            {
+                var ids = rows.Select(e => e.EventId).ToList();
+                await _siteQueue.MarkReconciledAsync(ids, ct).ConfigureAwait(false);
+            }
+
+            return new PullAuditEventsResponse(rows, MoreAvailable: rows.Count >= batchSize);
+        }
+    }
+
+    /// <summary>
+    /// In-memory enumerator returning a fixed single-site list — mirrors the
+    /// pattern used in <c>SiteAuditReconciliationActorTests</c>.
+    /// </summary>
+    private sealed class StaticEnumerator : ISiteEnumerator
+    {
+        private readonly IReadOnlyList<SiteEntry> _sites;
+        public StaticEnumerator(params SiteEntry[] sites) => _sites = sites;
+        public Task<IReadOnlyList<SiteEntry>> EnumerateAsync(CancellationToken ct = default) =>
+            Task.FromResult(_sites);
+    }
+
+    private ScadaLinkDbContext CreateContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    private static AuditEvent NewEvent(string siteId, DateTime occurredAt) => new()
+    {
+        EventId = Guid.NewGuid(),
+        OccurredAtUtc = occurredAt,
+        Channel = AuditChannel.ApiOutbound,
+        Kind = AuditKind.ApiCall,
+        Status = AuditStatus.Delivered,
+        SourceSiteId = siteId,
+        Target = "external-system-a/method",
+    };
+
+    private SqliteAuditWriter CreateInMemorySqliteWriter() =>
+        new SqliteAuditWriter(
+            Options.Create(new SqliteAuditWriterOptions
+            {
+                DatabasePath = "ignored",
+                BatchSize = 64,
+                ChannelCapacity = 4096,
+            }),
+            NullLogger<SqliteAuditWriter>.Instance,
+            connectionStringOverride:
+                $"Data Source=file:outage-{Guid.NewGuid():N}?mode=memory&cache=shared");
+
+    private (IServiceProvider Sp, IActorRef Ingest) BuildCentralPipeline()
+    {
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(opts =>
+            opts.UseSqlServer(_fixture.ConnectionString));
+        services.AddScoped<IAuditLogRepository>(sp =>
+            new AuditLogRepository(sp.GetRequiredService<ScadaLinkDbContext>()));
+        var sp = services.BuildServiceProvider();
+
+        var ingest = Sys.ActorOf(Props.Create(() => new AuditLogIngestActor(
+            sp,
+            NullLogger<AuditLogIngestActor>.Instance)));
+        return (sp, ingest);
+    }
+
+    private static SiteAuditReconciliationOptions FastTickOptions(int batchSize = 256) => new()
+    {
+        ReconciliationIntervalSeconds = 300,
+        ReconciliationIntervalOverride = TimeSpan.FromMilliseconds(100),
+        BatchSize = batchSize,
+        StalledAfterNonDrainingCycles = 2,
+    };
+
+    // ---------------------------------------------------------------------
+    // 1. CentralOutage_200Events_Buffer_Then_Reconciliation_Catches_Up_NoDuplicates
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task CentralOutage_200Events_Buffer_Then_Reconciliation_Catches_Up_NoDuplicates()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = "outage-recon-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+
+        // Step 1: site accumulates 200 audit events during the simulated
+        // central outage. The push path is NOT wired here — every row stays
+        // Pending in the site SQLite store until reconciliation runs.
+        await using var sqliteWriter = CreateInMemorySqliteWriter();
+        var baseOccurred = new DateTime(2026, 5, 20, 12, 0, 0, DateTimeKind.Utc);
+        const int totalEvents = 200;
+        var written = new List<AuditEvent>(totalEvents);
+
+        for (int i = 0; i < totalEvents; i++)
+        {
+            // Strictly monotonic OccurredAtUtc so the cursor can advance
+            // deterministically batch-by-batch — mirrors how a real script
+            // workload generates timestamps in wall-clock order.
+            var evt = NewEvent(siteId, baseOccurred.AddMilliseconds(i));
+            written.Add(evt);
+            await sqliteWriter.WriteAsync(evt);
+        }
+
+        // Sanity: every row is Pending (no push path wired, so nothing has
+        // been Forwarded or Reconciled yet).
+        var pending = await sqliteWriter.ReadPendingAsync(totalEvents + 10);
+        Assert.Equal(totalEvents, pending.Count);
+
+        // Step 2: central comes online — wire the ingest actor + reconciliation
+        // actor. The pull client wraps the site queue directly (the production
+        // shape is one RPC call); each pull advances the actor's cursor and
+        // flips rows on the site to Reconciled.
+        var (sp, ingest) = BuildCentralPipeline();
+        await using (sp as IAsyncDisposable ?? throw new InvalidOperationException())
+        {
+            var pullClient = new QueueBackedPullClient(sqliteWriter);
+            var enumerator = new StaticEnumerator(new SiteEntry(siteId, "http://test:8083"));
+
+            // BatchSize = 64 so the actor needs ~4 ticks to drain 200 rows.
+            // The "after 5 minutes" wording in the brief is satisfied by the
+            // fast-tick override (100 ms per tick) plus AwaitAssert giving
+            // the actor up to ~30 seconds to settle in real time.
+            var opts = FastTickOptions(batchSize: 64);
+
+            // Standalone DI scope for the reconciliation actor (it shares the
+            // ingest actor's IServiceProvider so both writers see the same
+            // EF context configuration).
+            var reconciliationActor = Sys.ActorOf(Props.Create(() => new SiteAuditReconciliationActor(
+                enumerator,
+                pullClient,
+                sp,
+                Options.Create(opts),
+                NullLogger<SiteAuditReconciliationActor>.Instance)));
+
+            // Step 3: assert central AuditLog has all 200 rows after the
+            // actor drains. Polling the real MSSQL repository — the test
+            // fixture has its own database so a count restricted to this
+            // SourceSiteId is exact.
+            await AwaitAssertAsync(async () =>
+                {
+                    await using var ctx = CreateContext();
+                    var count = await ctx.Set<AuditEvent>()
+                        .Where(e => e.SourceSiteId == siteId)
+                        .CountAsync();
+                    Assert.Equal(totalEvents, count);
+                },
+                duration: TimeSpan.FromSeconds(30),
+                interval: TimeSpan.FromMilliseconds(200));
+
+            // Step 4: assert site rows flipped to Reconciled.
+            // ReadPendingAsync only returns Pending rows; after a full drain
+            // it must be empty.
+            await AwaitAssertAsync(async () =>
+                {
+                    var stillPending = await sqliteWriter.ReadPendingAsync(totalEvents + 10);
+                    Assert.Empty(stillPending);
+                },
+                duration: TimeSpan.FromSeconds(10),
+                interval: TimeSpan.FromMilliseconds(100));
+
+            // Step 5: assert no duplicates by EventId — central must have
+            // exactly the 200 rows we wrote at the site (one row per EventId).
+            await using var verify = CreateContext();
+            var centralIds = await verify.Set<AuditEvent>()
+                .Where(e => e.SourceSiteId == siteId)
+                .Select(e => e.EventId)
+                .ToListAsync();
+            Assert.Equal(totalEvents, centralIds.Count);
+            Assert.Equal(totalEvents, centralIds.Distinct().Count());
+            // And every EventId we wrote at the site is present centrally.
+            Assert.True(written.All(w => centralIds.Contains(w.EventId)),
+                "every site-written EventId should be present centrally.");
+
+            // Tear the actor down before disposing the harness; the actor's
+            // PostStop cancels its scheduled timer.
+            Sys.Stop(reconciliationActor);
+        }
+    }
+
+    // ---------------------------------------------------------------------
+    // 2. ReconciliationPull_Idempotent_Across_Two_Cycles
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task ReconciliationPull_Idempotent_Across_Two_Cycles()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = "outage-idem-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        const int totalEvents = 50;
+
+        await using var sqliteWriter = CreateInMemorySqliteWriter();
+        var baseOccurred = new DateTime(2026, 5, 20, 13, 0, 0, DateTimeKind.Utc);
+        for (int i = 0; i < totalEvents; i++)
+        {
+            await sqliteWriter.WriteAsync(NewEvent(siteId, baseOccurred.AddMilliseconds(i)));
+        }
+
+        var (sp, _) = BuildCentralPipeline();
+        await using (sp as IAsyncDisposable ?? throw new InvalidOperationException())
+        {
+            var pullClient = new QueueBackedPullClient(sqliteWriter);
+            var enumerator = new StaticEnumerator(new SiteEntry(siteId, "http://test:8083"));
+
+            var reconciliationActor = Sys.ActorOf(Props.Create(() => new SiteAuditReconciliationActor(
+                enumerator,
+                pullClient,
+                sp,
+                Options.Create(FastTickOptions()),
+                NullLogger<SiteAuditReconciliationActor>.Instance)));
+
+            // Wait for the first drain cycle to complete.
+            await AwaitAssertAsync(async () =>
+                {
+                    await using var ctx = CreateContext();
+                    var count = await ctx.Set<AuditEvent>()
+                        .Where(e => e.SourceSiteId == siteId)
+                        .CountAsync();
+                    Assert.Equal(totalEvents, count);
+                },
+                duration: TimeSpan.FromSeconds(30),
+                interval: TimeSpan.FromMilliseconds(200));
+
+            // Wait for additional pull cycles to fire — the actor ticks every
+            // 100 ms so a 1 s settle leaves the actor with at least ~5 ticks
+            // past the initial drain. Each subsequent tick must be a no-op
+            // because every row is now Reconciled and outside the
+            // ReadPendingSinceAsync filter.
+            var callsAfterDrain = pullClient.CallCount;
+            await Task.Delay(TimeSpan.FromMilliseconds(800));
+            Assert.True(pullClient.CallCount > callsAfterDrain,
+                $"expected additional pull calls after drain to validate idempotency, got {pullClient.CallCount} after {callsAfterDrain}");
+
+            // Central count must still be exactly totalEvents — no duplicates
+            // even though the cursor + read-Reconciled-too semantics could
+            // theoretically re-fetch on the second cycle.
+            await using var verify = CreateContext();
+            var rows = await verify.Set<AuditEvent>()
+                .Where(e => e.SourceSiteId == siteId)
+                .ToListAsync();
+            Assert.Equal(totalEvents, rows.Count);
+            Assert.Equal(totalEvents, rows.Select(r => r.EventId).Distinct().Count());
+
+            Sys.Stop(reconciliationActor);
+        }
+    }
+}

tests/ScadaLink.AuditLog.Tests/Integration/PartitionMaintenanceTests.cs

@@ -0,0 +1,278 @@
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.Commons.Interfaces;
+using ScadaLink.ConfigurationDatabase;
+using ScadaLink.ConfigurationDatabase.Maintenance;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+
+namespace ScadaLink.AuditLog.Tests.Integration;
+
+/// <summary>
+/// Bundle F (#23 M6-T12) end-to-end tests for the
+/// <see cref="AuditLogPartitionMaintenanceService"/> hosted service running
+/// the real EF/MSSQL <see cref="AuditLogPartitionMaintenance"/> against the
+/// per-class <see cref="MsSqlMigrationFixture"/>. The migration seeds
+/// boundaries for every month Jan 2026 – Dec 2027, so the eager startup tick
+/// can be exercised both for the "future covered" no-op case and for the
+/// "lookahead larger than covered" SPLIT case.
+/// </summary>
+/// <remarks>
+/// Tests within this class share one fixture DB — boundaries added by one
+/// test persist across the next. Each test reads the max boundary at the
+/// start and computes its lookahead relative to it, mirroring the pattern
+/// used by the per-component <c>AuditLogPartitionMaintenanceTests</c> in
+/// <c>ScadaLink.ConfigurationDatabase.Tests</c>.
+/// </remarks>
+public class PartitionMaintenanceTests : IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public PartitionMaintenanceTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    private ScadaLinkDbContext CreateContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    /// <summary>
+    /// Builds the central-side DI graph for the hosted service: scoped EF
+    /// context + scoped <see cref="IPartitionMaintenance"/> matching how
+    /// <c>AddConfigurationDatabase</c> wires the production composition root.
+    /// </summary>
+    private ServiceProvider BuildProvider()
+    {
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(
+            opts => opts.UseSqlServer(_fixture.ConnectionString),
+            ServiceLifetime.Scoped);
+        services.AddScoped<IPartitionMaintenance, AuditLogPartitionMaintenance>();
+        return services.BuildServiceProvider();
+    }
+
+    private static async Task<DateTime?> ReadMaxBoundaryAsync(IServiceProvider sp)
+    {
+        await using var scope = sp.CreateAsyncScope();
+        var maintenance = scope.ServiceProvider.GetRequiredService<IPartitionMaintenance>();
+        return await maintenance.GetMaxBoundaryAsync();
+    }
+
+    /// <summary>
+    /// Mirrors the helper in
+    /// <c>AuditLogPartitionMaintenanceTests.LookaheadForExtraBoundaries</c>:
+    /// the smallest lookahead value that lands the SPLIT horizon exactly
+    /// <paramref name="extraBoundaries"/> months past the current max.
+    /// </summary>
+    private static int LookaheadForExtraBoundaries(DateTime max, int extraBoundaries)
+    {
+        var nowFirstOfNextMonth = FirstOfNextMonth(DateTime.UtcNow);
+        var monthsToMax = ((max.Year - nowFirstOfNextMonth.Year) * 12)
+            + max.Month - nowFirstOfNextMonth.Month;
+        return monthsToMax + extraBoundaries;
+    }
+
+    private static int LookaheadInsideExistingRange(DateTime max)
+    {
+        var now = DateTime.UtcNow;
+        var months = ((max.Year - now.Year) * 12) + max.Month - now.Month - 1;
+        return Math.Max(1, months);
+    }
+
+    private static DateTime FirstOfNextMonth(DateTime instant)
+    {
+        var firstOfThisMonth = new DateTime(instant.Year, instant.Month, 1, 0, 0, 0, DateTimeKind.Utc);
+        return firstOfThisMonth.AddMonths(1);
+    }
+
+    /// <summary>
+    /// Awaits one full tick of the hosted service. The service runs an
+    /// eager startup tick inside <see cref="AuditLogPartitionMaintenanceService.StartAsync"/>'s
+    /// continuation, but the continuation is dispatched on a background
+    /// Task.Run — so we poll the side effect (the boundary count or
+    /// max-boundary value) until it changes.
+    /// </summary>
+    private async Task StartAndAwaitStartupTickAsync(
+        AuditLogPartitionMaintenanceService svc,
+        Func<Task<bool>> awaitCondition,
+        TimeSpan timeout)
+    {
+        await svc.StartAsync(CancellationToken.None);
+        var deadline = DateTime.UtcNow + timeout;
+        while (DateTime.UtcNow < deadline)
+        {
+            if (await awaitCondition())
+            {
+                return;
+            }
+            await Task.Delay(50);
+        }
+    }
+
+    // ---------------------------------------------------------------------
+    // 1. EndToEnd_DefaultLookahead_NoSplit_WhenFutureCovered
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_DefaultLookahead_NoSplit_WhenFutureCovered()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var sp = BuildProvider();
+
+        // The migration seeds boundaries through Dec 2027. With default
+        // lookahead = 1 and today = ~2026-05-20, horizon =
+        // NormalizeToFirstOfMonth(now) + 1 = 2026-07-01, well within the
+        // seeded range, so the startup tick should issue zero SPLITs.
+        var maxBefore = await ReadMaxBoundaryAsync(sp);
+        Assert.NotNull(maxBefore);
+
+        // Skip if the fixture DB already has boundaries past Dec 2027 from
+        // a prior test in this class — the lookahead-already-covered path
+        // is what we want to exercise, regardless of how far past Dec 2027
+        // the boundary may be.
+        var opts = Options.Create(new AuditLogPartitionMaintenanceOptions
+        {
+            IntervalSeconds = 60, // long enough that only the startup tick fires inside the test window
+            LookaheadMonths = 1,
+        });
+
+        var svc = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            NullLogger<AuditLogPartitionMaintenanceService>.Instance);
+
+        // Drive the startup tick. There is no public completion handle;
+        // poll until either (a) the max boundary changes (which would be a
+        // failure for this test) or (b) the polling window expires (success).
+        await svc.StartAsync(CancellationToken.None);
+        await Task.Delay(TimeSpan.FromSeconds(2));
+        await svc.StopAsync(CancellationToken.None);
+        svc.Dispose();
+
+        // Assert the max boundary is unchanged: no SPLIT was issued.
+        var maxAfter = await ReadMaxBoundaryAsync(sp);
+        Assert.Equal(maxBefore, maxAfter);
+    }
+
+    // ---------------------------------------------------------------------
+    // 2. EndToEnd_LookaheadLargerThanCovered_Splits_NewBoundaries
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_LookaheadLargerThanCovered_Splits_NewBoundaries()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var sp = BuildProvider();
+
+        var maxBefore = await ReadMaxBoundaryAsync(sp);
+        Assert.NotNull(maxBefore);
+
+        // Pick a lookahead that adds exactly two new boundaries past the
+        // current max. The expected new boundaries are max+1mo and max+2mo.
+        var lookahead = LookaheadForExtraBoundaries(maxBefore.Value, extraBoundaries: 2);
+        var expectedFirstNew = maxBefore.Value.AddMonths(1);
+        var expectedSecondNew = maxBefore.Value.AddMonths(2);
+
+        var opts = Options.Create(new AuditLogPartitionMaintenanceOptions
+        {
+            IntervalSeconds = 60,
+            LookaheadMonths = lookahead,
+        });
+
+        var svc = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            NullLogger<AuditLogPartitionMaintenanceService>.Instance);
+
+        // Drive the startup tick. Wait until max boundary moves forward by
+        // the expected amount; SPLIT against MSSQL can take a second or two
+        // on a busy dev container.
+        await StartAndAwaitStartupTickAsync(
+            svc,
+            async () =>
+            {
+                var current = await ReadMaxBoundaryAsync(sp);
+                return current == expectedSecondNew;
+            },
+            timeout: TimeSpan.FromSeconds(15));
+
+        await svc.StopAsync(CancellationToken.None);
+        svc.Dispose();
+
+        var maxAfter = await ReadMaxBoundaryAsync(sp);
+        // Two new boundaries should be present after the startup tick. The
+        // hosted service does not surface the added-list directly (it logs
+        // only at Information), so we assert via the max-boundary delta.
+        Assert.Equal(expectedSecondNew, maxAfter);
+        // Sanity: the intermediate boundary was also added (the loop
+        // SPLITs every month from max+1 up to horizon, in order).
+        Assert.True(expectedFirstNew < expectedSecondNew);
+    }
+
+    // ---------------------------------------------------------------------
+    // 3. EndToEnd_PartitionMaintenance_Idempotent_OverTwoRuns
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_PartitionMaintenance_Idempotent_OverTwoRuns()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var sp = BuildProvider();
+
+        var maxBefore = await ReadMaxBoundaryAsync(sp);
+        Assert.NotNull(maxBefore);
+
+        // Add exactly one new boundary on the first run.
+        var lookahead = LookaheadForExtraBoundaries(maxBefore.Value, extraBoundaries: 1);
+        var expectedAdded = maxBefore.Value.AddMonths(1);
+
+        var opts = Options.Create(new AuditLogPartitionMaintenanceOptions
+        {
+            IntervalSeconds = 60,
+            LookaheadMonths = lookahead,
+        });
+
+        // First run.
+        var svc1 = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            NullLogger<AuditLogPartitionMaintenanceService>.Instance);
+        await StartAndAwaitStartupTickAsync(
+            svc1,
+            async () =>
+            {
+                var current = await ReadMaxBoundaryAsync(sp);
+                return current == expectedAdded;
+            },
+            timeout: TimeSpan.FromSeconds(15));
+        await svc1.StopAsync(CancellationToken.None);
+        svc1.Dispose();
+
+        var maxAfterFirst = await ReadMaxBoundaryAsync(sp);
+        Assert.Equal(expectedAdded, maxAfterFirst);
+
+        // Second run with the SAME lookahead value. Because the boundary
+        // is already covered, the EnsureLookaheadAsync call must be a
+        // no-op — max boundary is unchanged AND no exception is thrown.
+        var svc2 = new AuditLogPartitionMaintenanceService(
+            sp.GetRequiredService<IServiceScopeFactory>(),
+            opts,
+            NullLogger<AuditLogPartitionMaintenanceService>.Instance);
+        await svc2.StartAsync(CancellationToken.None);
+        // Wait long enough that the startup tick would have fired and
+        // logged any boundary addition; the boundary state must remain
+        // unchanged after the wait.
+        await Task.Delay(TimeSpan.FromSeconds(2));
+        await svc2.StopAsync(CancellationToken.None);
+        svc2.Dispose();
+
+        var maxAfterSecond = await ReadMaxBoundaryAsync(sp);
+        Assert.Equal(maxAfterFirst, maxAfterSecond);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Integration/PartitionPurgeTests.cs

@@ -0,0 +1,354 @@
+using Akka.Actor;
+using Akka.TestKit.Xunit2;
+using Microsoft.Data.SqlClient;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.DependencyInjection;
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Central;
+using ScadaLink.AuditLog.Configuration;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Types.Enums;
+using ScadaLink.ConfigurationDatabase;
+using ScadaLink.ConfigurationDatabase.Repositories;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+
+namespace ScadaLink.AuditLog.Tests.Integration;
+
+/// <summary>
+/// Bundle F (#23 M6-T11) end-to-end test for the daily partition-switch
+/// purge: seeds three monthly partitions (Jan / Feb / Mar 2026) with direct
+/// INSERTs that bypass the standard repository ingest path (so the seed
+/// timestamps are explicit), drives <see cref="AuditLogPurgeActor"/> against
+/// the real <see cref="AuditLogRepository"/> + per-test
+/// <see cref="MsSqlMigrationFixture"/> database, and asserts:
+/// <list type="number">
+///   <item>The oldest partition (Jan) is removed.</item>
+///   <item>Newer partitions (Feb + Mar) are untouched.</item>
+///   <item>The <c>UX_AuditLog_EventId</c> unique index survives the
+///         drop-and-rebuild dance.</item>
+///   <item><see cref="IAuditLogRepository.InsertIfNotExistsAsync"/> remains
+///         idempotent against the rebuilt index after the purge.</item>
+/// </list>
+/// </summary>
+/// <remarks>
+/// The brief calls out that direct INSERTs bypass the writer role's INSERT-only
+/// grant; the fixture connects as <c>sa</c> (see
+/// <see cref="MsSqlMigrationFixture"/>'s default admin connection string), so
+/// the seed step does not need the writer role at all. The drop-and-rebuild
+/// dance itself runs under the same admin connection because the test owns
+/// the database — the role granularity is exercised in the repository tests,
+/// not here.
+/// </remarks>
+public class PartitionPurgeTests : TestKit, IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public PartitionPurgeTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    private ScadaLinkDbContext CreateContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    /// <summary>
+    /// Direct INSERT into <c>dbo.AuditLog</c> bypassing
+    /// <see cref="IAuditLogRepository.InsertIfNotExistsAsync"/>. Used by the
+    /// seed step so the test can place rows in arbitrary partitions without
+    /// the repository's idempotency wrapper or ingest-stamping behaviour
+    /// affecting the seed payload.
+    /// </summary>
+    private async Task DirectInsertAsync(
+        SqlConnection conn,
+        Guid eventId,
+        DateTime occurredAtUtc,
+        string siteId)
+    {
+        await using var cmd = conn.CreateCommand();
+        cmd.CommandText = @"
+INSERT INTO dbo.AuditLog
+    (EventId, OccurredAtUtc, IngestedAtUtc, Channel, Kind, CorrelationId,
+     SourceSiteId, SourceInstanceId, SourceScript, Actor, Target, Status,
+     HttpStatus, DurationMs, ErrorMessage, ErrorDetail, RequestSummary,
+     ResponseSummary, PayloadTruncated, Extra, ForwardState)
+VALUES
+    (@EventId, @OccurredAtUtc, @IngestedAtUtc, 'ApiOutbound', 'ApiCall', NULL,
+     @SourceSiteId, NULL, NULL, NULL, NULL, 'Delivered',
+     NULL, NULL, NULL, NULL, NULL,
+     NULL, 0, NULL, NULL);";
+        cmd.Parameters.Add("@EventId", System.Data.SqlDbType.UniqueIdentifier).Value = eventId;
+        // SqlDbType.DateTime2 with explicit Scale 7 matches the
+        // OccurredAtUtc column shape (datetime2(7)) and avoids the implicit
+        // narrowing that SqlClient's default DateTime → datetime applies via
+        // AddWithValue. Critical for partition assignment: the partition
+        // function key column is datetime2(7); a narrowed value would still
+        // land in the correct partition for first-of-month seeds, but
+        // explicit typing here documents the intent and matches how the
+        // production repository INSERT shapes its parameters.
+        var occurredParam = cmd.Parameters.Add("@OccurredAtUtc", System.Data.SqlDbType.DateTime2);
+        occurredParam.Scale = 7;
+        occurredParam.Value = occurredAtUtc;
+        var ingestedParam = cmd.Parameters.Add("@IngestedAtUtc", System.Data.SqlDbType.DateTime2);
+        ingestedParam.Scale = 7;
+        ingestedParam.Value = DateTime.UtcNow;
+        cmd.Parameters.Add("@SourceSiteId", System.Data.SqlDbType.VarChar, 64).Value = siteId;
+        await cmd.ExecuteNonQueryAsync();
+    }
+
+    /// <summary>
+    /// Asserts that <c>UX_AuditLog_EventId</c> exists in
+    /// <c>sys.indexes</c>. The drop-and-rebuild dance briefly removes the
+    /// index inside its transaction; this check is meant to fire AFTER the
+    /// actor's purge tick has committed so the rebuilt index is observable.
+    /// </summary>
+    private static async Task AssertUxIndexExistsAsync(SqlConnection conn)
+    {
+        await using var cmd = conn.CreateCommand();
+        cmd.CommandText = @"
+SELECT COUNT(*)
+FROM   sys.indexes
+WHERE  name = 'UX_AuditLog_EventId'
+  AND  object_id = OBJECT_ID('dbo.AuditLog');";
+        var raw = await cmd.ExecuteScalarAsync();
+        var count = Convert.ToInt32(raw);
+        Assert.True(count == 1, $"UX_AuditLog_EventId should be present post-purge; sys.indexes count was {count}.");
+    }
+
+    private IActorRef CreateActor(
+        IServiceProvider sp,
+        AuditLogPurgeOptions purgeOptions,
+        AuditLogOptions auditOptions)
+    {
+        return Sys.ActorOf(Props.Create(() => new AuditLogPurgeActor(
+            sp,
+            Options.Create(purgeOptions),
+            Options.Create(auditOptions),
+            NullLogger<AuditLogPurgeActor>.Instance)));
+    }
+
+    private static (DateTime Jan, DateTime Feb, DateTime Mar) SeedOccurredAt() => (
+        new DateTime(2026, 1, 15, 0, 0, 0, DateTimeKind.Utc),
+        new DateTime(2026, 2, 15, 0, 0, 0, DateTimeKind.Utc),
+        new DateTime(2026, 3, 15, 0, 0, 0, DateTimeKind.Utc));
+
+    // ---------------------------------------------------------------------
+    // 1. EndToEnd_OldestPartition_PurgedViaActor_NewerKept
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_OldestPartition_PurgedViaActor_NewerKept()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        // Test date is ~2026-05-20 per environment. We want a threshold that
+        // sits strictly between Jan 15 (the Jan partition's MAX) and Feb 15
+        // (the Feb partition's MAX) so only the Jan-2026 partition is
+        // eligible for purge. RetentionDays = 100 gives a threshold of
+        // ~2026-02-09 — Jan 15 is older (purged), Feb 15 and Mar 15 are
+        // newer (kept). The window between Jan 15 and Feb 15 is wide enough
+        // (~30 days) to tolerate any plausible test-clock drift in CI.
+        var siteId = "purge-e2e-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var janEventId = Guid.NewGuid();
+        var febEventId = Guid.NewGuid();
+        var marEventId = Guid.NewGuid();
+        var (janOccurred, febOccurred, marOccurred) = SeedOccurredAt();
+
+        await using (var seedConn = _fixture.OpenConnection())
+        {
+            await DirectInsertAsync(seedConn, janEventId, janOccurred, siteId);
+            await DirectInsertAsync(seedConn, febEventId, febOccurred, siteId);
+            await DirectInsertAsync(seedConn, marEventId, marOccurred, siteId);
+        }
+
+        // Wire the actor with a real EF context against the fixture DB.
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(
+            opts => opts.UseSqlServer(_fixture.ConnectionString),
+            ServiceLifetime.Scoped);
+        services.AddScoped<IAuditLogRepository, AuditLogRepository>();
+        var sp = services.BuildServiceProvider();
+
+        var probe = CreateTestProbe();
+        Sys.EventStream.Subscribe(probe.Ref, typeof(AuditLogPurgedEvent));
+
+        var purgeOptions = new AuditLogPurgeOptions
+        {
+            IntervalHours = 24,
+            IntervalOverride = TimeSpan.FromMilliseconds(100),
+        };
+        var auditOptions = new AuditLogOptions { RetentionDays = 100 };
+
+        CreateActor(sp, purgeOptions, auditOptions);
+
+        // Wait for the actor's tick to purge the Jan-2026 partition.
+        // Concurrent test runs against the same fixture might also create
+        // eligible partitions, but each test class owns its own fixture DB
+        // (MsSqlMigrationFixture seeds a guid-named DB per class), so the
+        // Jan-2026 boundary is the only one this test can have produced.
+        var janBoundary = new DateTime(2026, 1, 1, 0, 0, 0, DateTimeKind.Utc);
+        var matched = probe.FishForMessage<AuditLogPurgedEvent>(
+            isMessage: m => m.MonthBoundary == janBoundary,
+            max: TimeSpan.FromSeconds(30));
+        Assert.True(matched.RowsDeleted >= 1,
+            $"Expected RowsDeleted >= 1 for Jan-2026 boundary; got {matched.RowsDeleted}.");
+
+        // Allow a brief settle in case the actor is mid-tick on Feb/Mar
+        // (it shouldn't be, since RetentionDays = 90 means only Jan is
+        // eligible, but the actor MAY re-enumerate quickly while we read).
+        await Task.Delay(TimeSpan.FromMilliseconds(500));
+
+        await using var verify = CreateContext();
+        var rows = await verify.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == siteId)
+            .ToListAsync();
+
+        // Jan removed; Feb + Mar untouched. Because the test owns the site
+        // id and the fixture DB, exact set membership is observable.
+        Assert.DoesNotContain(rows, r => r.EventId == janEventId);
+        Assert.Contains(rows, r => r.EventId == febEventId);
+        Assert.Contains(rows, r => r.EventId == marEventId);
+    }
+
+    // ---------------------------------------------------------------------
+    // 2. EndToEnd_UxIndexRebuilt_AfterPurge
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_UxIndexRebuilt_AfterPurge()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        // Same shape as test 1 — purge the Jan-2026 partition and then
+        // assert the UX_AuditLog_EventId index is still present. The
+        // drop-and-rebuild dance briefly removes it inside its transaction
+        // (the SWITCH PARTITION step requires the non-aligned unique index
+        // to be absent), but step 5 rebuilds it before committing. Sanity-
+        // checking the post-COMMIT shape here documents the invariant in an
+        // assertable way.
+        var siteId = "purge-uxidx-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var janEventId = Guid.NewGuid();
+        var (janOccurred, _, _) = SeedOccurredAt();
+
+        await using (var seedConn = _fixture.OpenConnection())
+        {
+            await DirectInsertAsync(seedConn, janEventId, janOccurred, siteId);
+        }
+
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(
+            opts => opts.UseSqlServer(_fixture.ConnectionString),
+            ServiceLifetime.Scoped);
+        services.AddScoped<IAuditLogRepository, AuditLogRepository>();
+        var sp = services.BuildServiceProvider();
+
+        var probe = CreateTestProbe();
+        Sys.EventStream.Subscribe(probe.Ref, typeof(AuditLogPurgedEvent));
+
+        CreateActor(
+            sp,
+            new AuditLogPurgeOptions
+            {
+                IntervalHours = 24,
+                IntervalOverride = TimeSpan.FromMilliseconds(100),
+            },
+            new AuditLogOptions { RetentionDays = 90 });
+
+        var janBoundary = new DateTime(2026, 1, 1, 0, 0, 0, DateTimeKind.Utc);
+        probe.FishForMessage<AuditLogPurgedEvent>(
+            isMessage: m => m.MonthBoundary == janBoundary,
+            max: TimeSpan.FromSeconds(30));
+
+        // Open a fresh connection (the actor's pool is owned by EF) and
+        // assert the index is present post-purge.
+        await using var check = _fixture.OpenConnection();
+        await AssertUxIndexExistsAsync(check);
+    }
+
+    // ---------------------------------------------------------------------
+    // 3. EndToEnd_InsertIfNotExistsAsync_StillIdempotent_AfterPurge
+    // ---------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task EndToEnd_InsertIfNotExistsAsync_StillIdempotent_AfterPurge()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        // Seed + purge a Jan-2026 row, THEN exercise InsertIfNotExistsAsync
+        // twice for a fresh (May-2026) EventId. The second call must be a
+        // no-op (duplicate-key collision swallowed by the repository, per
+        // M2 Bundle A's race-fix) — which means the rebuilt
+        // UX_AuditLog_EventId unique index is functioning as intended.
+        var siteId = "purge-idem-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var janEventId = Guid.NewGuid();
+        var (janOccurred, _, _) = SeedOccurredAt();
+
+        await using (var seedConn = _fixture.OpenConnection())
+        {
+            await DirectInsertAsync(seedConn, janEventId, janOccurred, siteId);
+        }
+
+        var services = new ServiceCollection();
+        services.AddDbContext<ScadaLinkDbContext>(
+            opts => opts.UseSqlServer(_fixture.ConnectionString),
+            ServiceLifetime.Scoped);
+        services.AddScoped<IAuditLogRepository, AuditLogRepository>();
+        var sp = services.BuildServiceProvider();
+
+        var probe = CreateTestProbe();
+        Sys.EventStream.Subscribe(probe.Ref, typeof(AuditLogPurgedEvent));
+
+        CreateActor(
+            sp,
+            new AuditLogPurgeOptions
+            {
+                IntervalHours = 24,
+                IntervalOverride = TimeSpan.FromMilliseconds(100),
+            },
+            new AuditLogOptions { RetentionDays = 90 });
+
+        var janBoundary = new DateTime(2026, 1, 1, 0, 0, 0, DateTimeKind.Utc);
+        probe.FishForMessage<AuditLogPurgedEvent>(
+            isMessage: m => m.MonthBoundary == janBoundary,
+            max: TimeSpan.FromSeconds(30));
+
+        // Settle then exercise InsertIfNotExistsAsync twice for the same
+        // EventId. The repository's idempotency relies on
+        // UX_AuditLog_EventId being present so the IF NOT EXISTS … INSERT
+        // race window resolves to a duplicate-key violation the repo
+        // swallows. If the index were missing here, two rows would land
+        // and the second InsertIfNotExistsAsync would silently double-insert.
+        await Task.Delay(TimeSpan.FromMilliseconds(500));
+
+        var freshEventId = Guid.NewGuid();
+        var freshOccurred = new DateTime(2026, 5, 15, 12, 0, 0, DateTimeKind.Utc);
+        var freshSite = "purge-idem-fresh-" + Guid.NewGuid().ToString("N").Substring(0, 8);
+        var freshEvt = new AuditEvent
+        {
+            EventId = freshEventId,
+            OccurredAtUtc = freshOccurred,
+            Channel = AuditChannel.ApiOutbound,
+            Kind = AuditKind.ApiCall,
+            Status = AuditStatus.Delivered,
+            SourceSiteId = freshSite,
+            Target = "system-x/method",
+        };
+
+        await using (var ctx = CreateContext())
+        {
+            var repo = new AuditLogRepository(ctx);
+            await repo.InsertIfNotExistsAsync(freshEvt);
+            // Same row a second time — must be a silent no-op.
+            await repo.InsertIfNotExistsAsync(freshEvt);
+        }
+
+        await using var verify = CreateContext();
+        var rows = await verify.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == freshSite)
+            .ToListAsync();
+        Assert.Single(rows);
+        Assert.Equal(freshEventId, rows[0].EventId);
+    }
+}

tests/ScadaLink.AuditLog.Tests/Integration/SyncCallEmissionEndToEndTests.cs

@@ -9,6 +9,7 @@ using ScadaLink.AuditLog.Site.Telemetry;
 using ScadaLink.AuditLog.Tests.Integration.Infrastructure;
 using ScadaLink.Commons.Entities.Audit;
 using ScadaLink.Commons.Interfaces.Repositories;
+using ScadaLink.Commons.Interfaces.Services;
 using ScadaLink.Commons.Types.Audit;
 using ScadaLink.Commons.Types.Enums;
 using ScadaLink.ConfigurationDatabase;

tests/ScadaLink.AuditLog.Tests/Site/SqliteAuditWriterBacklogStatsTests.cs

@@ -0,0 +1,136 @@
+using Microsoft.Extensions.Logging.Abstractions;
+using Microsoft.Extensions.Options;
+using ScadaLink.AuditLog.Site;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Types.Enums;
+
+namespace ScadaLink.AuditLog.Tests.Site;
+
+/// <summary>
+/// Bundle E (M6-T6) tests for <see cref="SqliteAuditWriter.GetBacklogStatsAsync"/>.
+/// Exercises the health-metric surface that <c>SiteAuditBacklogReporter</c>
+/// polls every 30 s and pushes onto the site health report as
+/// <c>SiteAuditBacklog</c>.
+/// </summary>
+public class SqliteAuditWriterBacklogStatsTests : IDisposable
+{
+    private readonly string _dbPath;
+
+    public SqliteAuditWriterBacklogStatsTests()
+    {
+        // OnDiskBytes assertions only make sense against a real file — the
+        // shared-cache in-memory mode returns 0 for the file size, so this
+        // suite is opinionated about file-backed storage. Tests in
+        // SqliteAuditWriterWriteTests use in-memory for performance reasons.
+        _dbPath = Path.Combine(Path.GetTempPath(),
+            $"audit-backlog-stats-{Guid.NewGuid():N}.db");
+    }
+
+    public void Dispose()
+    {
+        if (File.Exists(_dbPath))
+        {
+            try { File.Delete(_dbPath); } catch { /* test cleanup best-effort */ }
+        }
+    }
+
+    private SqliteAuditWriter CreateWriter()
+    {
+        var options = new SqliteAuditWriterOptions { DatabasePath = _dbPath };
+        return new SqliteAuditWriter(
+            Options.Create(options),
+            NullLogger<SqliteAuditWriter>.Instance);
+    }
+
+    private static AuditEvent NewEvent(DateTime? occurredAtUtc = null) => new()
+    {
+        EventId = Guid.NewGuid(),
+        OccurredAtUtc = occurredAtUtc ?? DateTime.UtcNow,
+        Channel = AuditChannel.ApiOutbound,
+        Kind = AuditKind.ApiCall,
+        Status = AuditStatus.Delivered,
+        PayloadTruncated = false,
+    };
+
+    [Fact]
+    public async Task EmptyDb_Returns_Zero_Null_AndZeroBytes()
+    {
+        // No file exists yet — the writer ctor creates one but no rows are
+        // inserted; the snapshot should report a clean queue. OnDiskBytes is
+        // allowed to be zero (fresh ftruncate) OR small (page header) — the
+        // contract only requires non-negative; we assert >= 0 and exercise
+        // the pending fields strictly.
+        await using var writer = CreateWriter();
+
+        var snapshot = await writer.GetBacklogStatsAsync();
+
+        Assert.Equal(0, snapshot.PendingCount);
+        Assert.Null(snapshot.OldestPendingUtc);
+        Assert.True(snapshot.OnDiskBytes >= 0,
+            $"OnDiskBytes must be non-negative, got {snapshot.OnDiskBytes}");
+    }
+
+    [Fact]
+    public async Task Pending_5_Returns_5()
+    {
+        await using var writer = CreateWriter();
+
+        for (var i = 0; i < 5; i++)
+        {
+            await writer.WriteAsync(NewEvent());
+        }
+
+        var snapshot = await writer.GetBacklogStatsAsync();
+
+        Assert.Equal(5, snapshot.PendingCount);
+    }
+
+    [Fact]
+    public async Task OldestPending_Is_Earliest_OccurredAtUtc()
+    {
+        await using var writer = CreateWriter();
+
+        var t1 = new DateTime(2026, 5, 20, 10, 0, 0, DateTimeKind.Utc);
+        var t2 = new DateTime(2026, 5, 20, 10, 1, 0, DateTimeKind.Utc);
+        var t3 = new DateTime(2026, 5, 20, 10, 2, 0, DateTimeKind.Utc);
+
+        // Insert out of order so the snapshot is not "the last write" by
+        // accident — the OldestPendingUtc must come from a column-min, not
+        // an insertion-order proxy.
+        await writer.WriteAsync(NewEvent(t2));
+        await writer.WriteAsync(NewEvent(t1));
+        await writer.WriteAsync(NewEvent(t3));
+
+        var snapshot = await writer.GetBacklogStatsAsync();
+
+        Assert.Equal(3, snapshot.PendingCount);
+        Assert.NotNull(snapshot.OldestPendingUtc);
+        // The DB round-trips OccurredAtUtc through the "o" format which
+        // preserves Kind=Utc — assert tick-equality.
+        Assert.Equal(t1, snapshot.OldestPendingUtc!.Value);
+    }
+
+    [Fact]
+    public async Task OnDiskBytes_ReturnsFileSize()
+    {
+        await using var writer = CreateWriter();
+
+        // Insert enough rows to grow the file past the empty schema baseline.
+        for (var i = 0; i < 100; i++)
+        {
+            await writer.WriteAsync(NewEvent());
+        }
+
+        var snapshot = await writer.GetBacklogStatsAsync();
+
+        // The exact size depends on SQLite page allocation, but a file-backed
+        // db with 100 inserted rows MUST be larger than the empty schema
+        // (a few pages, ~4 KB). The implementation should return the
+        // FileInfo.Length value verbatim.
+        Assert.True(File.Exists(_dbPath), $"DB file should exist at {_dbPath}");
+        var expected = new FileInfo(_dbPath).Length;
+        Assert.Equal(expected, snapshot.OnDiskBytes);
+        Assert.True(snapshot.OnDiskBytes > 0,
+            $"after 100 inserts OnDiskBytes must be > 0, got {snapshot.OnDiskBytes}");
+    }
+}

tests/ScadaLink.AuditLog.Tests/Site/SqliteAuditWriterWriteTests.cs

@@ -204,4 +204,153 @@ public class SqliteAuditWriterWriteTests
         await writer.MarkForwardedAsync(phantomIds);
         // No assertion needed: the call must complete without throwing.
     }
+
+    // ----- M6 reconciliation pull surface ----- //
+
+    [Fact]
+    public async Task ReadPendingSinceAsync_Returns_PendingAndForwarded_OldestFirst_LimitedToN()
+    {
+        var (writer, dataSource) = CreateWriter(nameof(ReadPendingSinceAsync_Returns_PendingAndForwarded_OldestFirst_LimitedToN));
+        await using var _ = writer;
+
+        var baseTime = new DateTime(2026, 5, 20, 12, 0, 0, DateTimeKind.Utc);
+        var evts = new[]
+        {
+            NewEvent(occurredAtUtc: baseTime.AddSeconds(5)),
+            NewEvent(occurredAtUtc: baseTime.AddSeconds(1)),
+            NewEvent(occurredAtUtc: baseTime.AddSeconds(3)),
+            NewEvent(occurredAtUtc: baseTime.AddSeconds(2)),
+            NewEvent(occurredAtUtc: baseTime.AddSeconds(4)),
+        };
+        foreach (var e in evts) await writer.WriteAsync(e);
+
+        // Flip half to Forwarded — they must still surface in the reconciliation pull
+        // because central hasn't confirmed they were ingested yet.
+        await writer.MarkForwardedAsync(new[] { evts[0].EventId, evts[2].EventId });
+
+        var rows = await writer.ReadPendingSinceAsync(sinceUtc: DateTime.MinValue, batchSize: 3);
+
+        Assert.Equal(3, rows.Count);
+        Assert.Equal(baseTime.AddSeconds(1), rows[0].OccurredAtUtc);
+        Assert.Equal(baseTime.AddSeconds(2), rows[1].OccurredAtUtc);
+        Assert.Equal(baseTime.AddSeconds(3), rows[2].OccurredAtUtc);
+    }
+
+    [Fact]
+    public async Task ReadPendingSinceAsync_ExcludesRowsOlderThanSinceUtc()
+    {
+        var (writer, _) = CreateWriter(nameof(ReadPendingSinceAsync_ExcludesRowsOlderThanSinceUtc));
+        await using var _w = writer;
+
+        var baseTime = new DateTime(2026, 5, 20, 12, 0, 0, DateTimeKind.Utc);
+        var old = NewEvent(occurredAtUtc: baseTime.AddSeconds(-30));
+        var newer1 = NewEvent(occurredAtUtc: baseTime.AddSeconds(10));
+        var newer2 = NewEvent(occurredAtUtc: baseTime.AddSeconds(20));
+
+        await writer.WriteAsync(old);
+        await writer.WriteAsync(newer1);
+        await writer.WriteAsync(newer2);
+
+        var rows = await writer.ReadPendingSinceAsync(sinceUtc: baseTime, batchSize: 10);
+
+        Assert.Equal(2, rows.Count);
+        Assert.Contains(rows, r => r.EventId == newer1.EventId);
+        Assert.Contains(rows, r => r.EventId == newer2.EventId);
+        Assert.DoesNotContain(rows, r => r.EventId == old.EventId);
+    }
+
+    [Fact]
+    public async Task ReadPendingSinceAsync_ExcludesReconciledRows()
+    {
+        var (writer, _) = CreateWriter(nameof(ReadPendingSinceAsync_ExcludesReconciledRows));
+        await using var _w = writer;
+
+        var baseTime = new DateTime(2026, 5, 20, 12, 0, 0, DateTimeKind.Utc);
+        var pending = NewEvent(occurredAtUtc: baseTime);
+        var reconciled = NewEvent(occurredAtUtc: baseTime.AddSeconds(1));
+
+        await writer.WriteAsync(pending);
+        await writer.WriteAsync(reconciled);
+        await writer.MarkReconciledAsync(new[] { reconciled.EventId });
+
+        var rows = await writer.ReadPendingSinceAsync(sinceUtc: DateTime.MinValue, batchSize: 10);
+
+        Assert.Single(rows);
+        Assert.Equal(pending.EventId, rows[0].EventId);
+    }
+
+    [Fact]
+    public async Task ReadPendingSinceAsync_InvalidBatchSize_Throws()
+    {
+        var (writer, _) = CreateWriter(nameof(ReadPendingSinceAsync_InvalidBatchSize_Throws));
+        await using var _w = writer;
+
+        await Assert.ThrowsAsync<ArgumentOutOfRangeException>(
+            () => writer.ReadPendingSinceAsync(DateTime.MinValue, batchSize: 0));
+        await Assert.ThrowsAsync<ArgumentOutOfRangeException>(
+            () => writer.ReadPendingSinceAsync(DateTime.MinValue, batchSize: -3));
+    }
+
+    [Fact]
+    public async Task MarkReconciledAsync_FlipsPendingAndForwarded_To_Reconciled()
+    {
+        var (writer, dataSource) = CreateWriter(nameof(MarkReconciledAsync_FlipsPendingAndForwarded_To_Reconciled));
+        await using var _ = writer;
+
+        var a = NewEvent();
+        var b = NewEvent();
+        var c = NewEvent();
+        await writer.WriteAsync(a);
+        await writer.WriteAsync(b);
+        await writer.WriteAsync(c);
+
+        // b is currently Forwarded; a and c are Pending.
+        await writer.MarkForwardedAsync(new[] { b.EventId });
+
+        await writer.MarkReconciledAsync(new[] { a.EventId, b.EventId, c.EventId });
+
+        using var connection = OpenVerifierConnection(dataSource);
+        using var cmd = connection.CreateCommand();
+        cmd.CommandText = "SELECT ForwardState, COUNT(*) FROM AuditLog GROUP BY ForwardState;";
+        using var reader = cmd.ExecuteReader();
+        var byState = new Dictionary<string, long>();
+        while (reader.Read())
+        {
+            byState[reader.GetString(0)] = reader.GetInt64(1);
+        }
+
+        Assert.Equal(3, byState[AuditForwardState.Reconciled.ToString()]);
+        Assert.False(byState.ContainsKey(AuditForwardState.Pending.ToString()));
+        Assert.False(byState.ContainsKey(AuditForwardState.Forwarded.ToString()));
+    }
+
+    [Fact]
+    public async Task MarkReconciledAsync_Idempotent_LeavesAlreadyReconciledRowsUntouched()
+    {
+        var (writer, dataSource) = CreateWriter(nameof(MarkReconciledAsync_Idempotent_LeavesAlreadyReconciledRowsUntouched));
+        await using var _ = writer;
+
+        var a = NewEvent();
+        await writer.WriteAsync(a);
+        await writer.MarkReconciledAsync(new[] { a.EventId });
+        // Re-call must not throw and must leave the single row Reconciled.
+        await writer.MarkReconciledAsync(new[] { a.EventId });
+
+        using var connection = OpenVerifierConnection(dataSource);
+        using var cmd = connection.CreateCommand();
+        cmd.CommandText = "SELECT ForwardState FROM AuditLog WHERE EventId = $id;";
+        cmd.Parameters.AddWithValue("$id", a.EventId.ToString());
+
+        Assert.Equal(AuditForwardState.Reconciled.ToString(), cmd.ExecuteScalar() as string);
+    }
+
+    [Fact]
+    public async Task MarkReconciledAsync_NonExistentId_NoThrow()
+    {
+        var (writer, _) = CreateWriter(nameof(MarkReconciledAsync_NonExistentId_NoThrow));
+        await using var _w = writer;
+
+        await writer.MarkReconciledAsync(new[] { Guid.NewGuid(), Guid.NewGuid() });
+        // Completes without throwing.
+    }
 }

tests/ScadaLink.AuditLog.Tests/Site/Telemetry/SiteAuditTelemetryActorTests.cs

@@ -7,6 +7,7 @@ using NSubstitute;
 using NSubstitute.ExceptionExtensions;
 using ScadaLink.AuditLog.Site.Telemetry;
 using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Services;
 using ScadaLink.Commons.Types.Enums;
 using ScadaLink.Communication.Grpc;
 

tests/ScadaLink.Communication.Tests/Protos/PullAuditEventsProtoTests.cs

@@ -0,0 +1,83 @@
+using Google.Protobuf;
+using Google.Protobuf.WellKnownTypes;
+using ScadaLink.Communication.Grpc;
+
+namespace ScadaLink.Communication.Tests.Protos;
+
+/// <summary>
+/// Wire-format round-trip tests for the Audit Log (#23) M6 reconciliation
+/// pull proto messages (<see cref="PullAuditEventsRequest"/>,
+/// <see cref="PullAuditEventsResponse"/>). Locks the additive contract the
+/// central→site reconciliation puller depends on.
+/// </summary>
+public class PullAuditEventsProtoTests
+{
+    private static AuditEventDto NewAuditDto(Guid? id = null) => new()
+    {
+        EventId = (id ?? Guid.NewGuid()).ToString(),
+        OccurredAtUtc = Timestamp.FromDateTimeOffset(
+            new DateTimeOffset(2026, 5, 20, 10, 15, 30, 123, TimeSpan.Zero)),
+        Channel = "ApiOutbound",
+        Kind = "ApiCall",
+        Status = "Delivered",
+        SourceSiteId = "site-1",
+    };
+
+    [Fact]
+    public void PullAuditEventsRequest_RoundTrip()
+    {
+        var sinceUtc = Timestamp.FromDateTimeOffset(
+            new DateTimeOffset(2026, 5, 20, 9, 0, 0, TimeSpan.Zero));
+
+        var original = new PullAuditEventsRequest
+        {
+            SinceUtc = sinceUtc,
+            BatchSize = 250,
+        };
+
+        var bytes = original.ToByteArray();
+        var deserialized = PullAuditEventsRequest.Parser.ParseFrom(bytes);
+
+        Assert.Equal(sinceUtc, deserialized.SinceUtc);
+        Assert.Equal(250, deserialized.BatchSize);
+    }
+
+    [Fact]
+    public void PullAuditEventsResponse_RoundTrip_WithEvents_And_MoreAvailable()
+    {
+        var dtos = Enumerable.Range(0, 4).Select(_ => NewAuditDto()).ToList();
+
+        var original = new PullAuditEventsResponse
+        {
+            MoreAvailable = true,
+        };
+        original.Events.AddRange(dtos);
+
+        var bytes = original.ToByteArray();
+        var deserialized = PullAuditEventsResponse.Parser.ParseFrom(bytes);
+
+        Assert.True(deserialized.MoreAvailable);
+        Assert.Equal(4, deserialized.Events.Count);
+        for (int i = 0; i < dtos.Count; i++)
+        {
+            Assert.Equal(dtos[i].EventId, deserialized.Events[i].EventId);
+            Assert.Equal(dtos[i].Status, deserialized.Events[i].Status);
+            Assert.Equal(dtos[i].SourceSiteId, deserialized.Events[i].SourceSiteId);
+            Assert.Equal(dtos[i].OccurredAtUtc, deserialized.Events[i].OccurredAtUtc);
+        }
+    }
+
+    [Fact]
+    public void PullAuditEventsResponse_Empty_Yields_EmptyEvents()
+    {
+        var original = new PullAuditEventsResponse();
+        Assert.Empty(original.Events);
+        Assert.False(original.MoreAvailable);
+
+        var bytes = original.ToByteArray();
+        var deserialized = PullAuditEventsResponse.Parser.ParseFrom(bytes);
+
+        Assert.Empty(deserialized.Events);
+        Assert.False(deserialized.MoreAvailable);
+    }
+}

tests/ScadaLink.Communication.Tests/SiteStreamPullAuditEventsTests.cs

@@ -0,0 +1,185 @@
+using Akka.TestKit.Xunit2;
+using Google.Protobuf.WellKnownTypes;
+using Grpc.Core;
+using Microsoft.Extensions.Logging.Abstractions;
+using NSubstitute;
+using NSubstitute.ExceptionExtensions;
+using ScadaLink.Commons.Entities.Audit;
+using ScadaLink.Commons.Interfaces.Services;
+using ScadaLink.Commons.Types.Enums;
+using ScadaLink.Communication.Grpc;
+
+namespace ScadaLink.Communication.Tests;
+
+/// <summary>
+/// Bundle A A2 tests for <see cref="SiteStreamGrpcServer.PullAuditEvents"/>.
+/// Verifies the request → ISiteAuditQueue.ReadPendingSinceAsync → response →
+/// MarkReconciledAsync round-trip through the gRPC handler. The queue is an
+/// NSubstitute stub so the tests never touch SQLite.
+/// </summary>
+public class SiteStreamPullAuditEventsTests : TestKit
+{
+    private readonly ISiteStreamSubscriber _subscriber = Substitute.For<ISiteStreamSubscriber>();
+
+    private SiteStreamGrpcServer CreateServer() =>
+        new(_subscriber, NullLogger<SiteStreamGrpcServer>.Instance);
+
+    private static ServerCallContext NewContext(CancellationToken ct = default)
+    {
+        var context = Substitute.For<ServerCallContext>();
+        context.CancellationToken.Returns(ct);
+        return context;
+    }
+
+    private static AuditEvent NewEvent(DateTime? occurredAt = null) => new()
+    {
+        EventId = Guid.NewGuid(),
+        OccurredAtUtc = occurredAt
+            ?? DateTime.SpecifyKind(new DateTime(2026, 5, 20, 10, 0, 0), DateTimeKind.Utc),
+        Channel = AuditChannel.ApiOutbound,
+        Kind = AuditKind.ApiCall,
+        Status = AuditStatus.Delivered,
+        SourceSiteId = "site-1",
+        PayloadTruncated = false,
+        ForwardState = AuditForwardState.Pending,
+    };
+
+    [Fact]
+    public async Task PullAuditEvents_NoQueueWired_ReturnsEmptyResponse()
+    {
+        var server = CreateServer();
+        // Intentionally do NOT call SetSiteAuditQueue — simulates a central-only
+        // host or a wiring-incomplete startup window.
+
+        var request = new PullAuditEventsRequest
+        {
+            SinceUtc = Timestamp.FromDateTime(DateTime.UtcNow.AddMinutes(-5)),
+            BatchSize = 100,
+        };
+
+        var response = await server.PullAuditEvents(request, NewContext());
+
+        Assert.Empty(response.Events);
+        Assert.False(response.MoreAvailable);
+    }
+
+    [Fact]
+    public async Task PullAuditEvents_With5PendingRows_ReturnsAllFiveDtos_AndFlipsToReconciled()
+    {
+        var queue = Substitute.For<ISiteAuditQueue>();
+        var events = Enumerable.Range(0, 5).Select(_ => NewEvent()).ToList();
+        queue.ReadPendingSinceAsync(Arg.Any<DateTime>(), Arg.Any<int>(), Arg.Any<CancellationToken>())
+            .Returns((IReadOnlyList<AuditEvent>)events);
+
+        var server = CreateServer();
+        server.SetSiteAuditQueue(queue);
+
+        var request = new PullAuditEventsRequest
+        {
+            SinceUtc = Timestamp.FromDateTime(DateTime.UtcNow.AddHours(-1)),
+            BatchSize = 100, // larger than returned count so MoreAvailable should be false
+        };
+
+        var response = await server.PullAuditEvents(request, NewContext());
+
+        Assert.Equal(5, response.Events.Count);
+        Assert.False(response.MoreAvailable); // 5 < 100
+        var expectedIds = events.Select(e => e.EventId.ToString()).ToHashSet();
+        Assert.True(expectedIds.SetEquals(response.Events.Select(d => d.EventId).ToHashSet()));
+
+        // Verify MarkReconciledAsync received the same 5 ids (best-effort flip).
+        await queue.Received(1).MarkReconciledAsync(
+            Arg.Is<IReadOnlyList<Guid>>(ids => ids.Count == 5 &&
+                                               ids.ToHashSet().SetEquals(events.Select(e => e.EventId))),
+            Arg.Any<CancellationToken>());
+    }
+
+    [Fact]
+    public async Task PullAuditEvents_RowsOlderThanSinceUtc_Excluded()
+    {
+        // The handler delegates the since-utc filter to ReadPendingSinceAsync;
+        // this test verifies it passes the request value through verbatim
+        // (no clock skew, no off-by-one) and that an empty queue response
+        // yields an empty gRPC response.
+        var queue = Substitute.For<ISiteAuditQueue>();
+        var capturedSince = DateTime.MinValue;
+        queue.ReadPendingSinceAsync(Arg.Any<DateTime>(), Arg.Any<int>(), Arg.Any<CancellationToken>())
+            .Returns(call =>
+            {
+                capturedSince = call.ArgAt<DateTime>(0);
+                return (IReadOnlyList<AuditEvent>)Array.Empty<AuditEvent>();
+            });
+
+        var server = CreateServer();
+        server.SetSiteAuditQueue(queue);
+
+        var since = DateTime.SpecifyKind(new DateTime(2026, 5, 20, 9, 30, 0), DateTimeKind.Utc);
+        var request = new PullAuditEventsRequest
+        {
+            SinceUtc = Timestamp.FromDateTime(since),
+            BatchSize = 50,
+        };
+
+        var response = await server.PullAuditEvents(request, NewContext());
+
+        Assert.Empty(response.Events);
+        Assert.False(response.MoreAvailable);
+        Assert.Equal(since, capturedSince);
+        // Empty result → no MarkReconciledAsync call (no rows to flip).
+        await queue.DidNotReceive().MarkReconciledAsync(
+            Arg.Any<IReadOnlyList<Guid>>(), Arg.Any<CancellationToken>());
+    }
+
+    [Fact]
+    public async Task PullAuditEvents_BatchSize3_Returns3Rows_MoreAvailableTrue()
+    {
+        var queue = Substitute.For<ISiteAuditQueue>();
+        var events = Enumerable.Range(0, 3).Select(_ => NewEvent()).ToList();
+        queue.ReadPendingSinceAsync(Arg.Any<DateTime>(), Arg.Any<int>(), Arg.Any<CancellationToken>())
+            .Returns((IReadOnlyList<AuditEvent>)events);
+
+        var server = CreateServer();
+        server.SetSiteAuditQueue(queue);
+
+        var request = new PullAuditEventsRequest
+        {
+            SinceUtc = Timestamp.FromDateTime(DateTime.UtcNow.AddHours(-1)),
+            BatchSize = 3,
+        };
+
+        var response = await server.PullAuditEvents(request, NewContext());
+
+        Assert.Equal(3, response.Events.Count);
+        // saturated batch → central needs to know to issue a follow-up pull
+        Assert.True(response.MoreAvailable);
+    }
+
+    [Fact]
+    public async Task PullAuditEvents_MarkReconciledThrows_ResponseStillReturned()
+    {
+        // The Reconciled flip is best-effort — if it fails, the response must
+        // still surface so central can ingest the rows (and dedup on EventId
+        // when it pulls them again).
+        var queue = Substitute.For<ISiteAuditQueue>();
+        var events = Enumerable.Range(0, 2).Select(_ => NewEvent()).ToList();
+        queue.ReadPendingSinceAsync(Arg.Any<DateTime>(), Arg.Any<int>(), Arg.Any<CancellationToken>())
+            .Returns((IReadOnlyList<AuditEvent>)events);
+        queue.MarkReconciledAsync(Arg.Any<IReadOnlyList<Guid>>(), Arg.Any<CancellationToken>())
+            .ThrowsAsync(new InvalidOperationException("SQLite disposed mid-call"));
+
+        var server = CreateServer();
+        server.SetSiteAuditQueue(queue);
+
+        var request = new PullAuditEventsRequest
+        {
+            SinceUtc = Timestamp.FromDateTime(DateTime.UtcNow.AddHours(-1)),
+            BatchSize = 100,
+        };
+
+        // Must NOT throw — the response is built before the flip and returned
+        // regardless of the flip outcome.
+        var response = await server.PullAuditEvents(request, NewContext());
+
+        Assert.Equal(2, response.Events.Count);
+    }
+}

tests/ScadaLink.ConfigurationDatabase.Tests/Maintenance/AuditLogPartitionMaintenanceTests.cs

@@ -0,0 +1,182 @@
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.Logging.Abstractions;
+using ScadaLink.ConfigurationDatabase.Maintenance;
+using ScadaLink.ConfigurationDatabase.Tests.Migrations;
+using Xunit;
+
+namespace ScadaLink.ConfigurationDatabase.Tests.Maintenance;
+
+/// <summary>
+/// Bundle D (#23 M6-T5) integration tests for
+/// <see cref="AuditLogPartitionMaintenance"/>. Uses the same
+/// <see cref="MsSqlMigrationFixture"/> as the AuditLog migration / repository
+/// tests so the ALTER PARTITION FUNCTION DDL runs against the actual seeded
+/// <c>pf_AuditLog_Month</c>.
+/// </summary>
+/// <remarks>
+/// The migration seeds boundaries for every month in 2026 and 2027 (Jan 2026
+/// through Dec 2027). Tests pick a lookahead relative to the current
+/// max-boundary at test start (rather than a fixed-target date) so each test
+/// is robust against earlier tests in the class having added boundaries to
+/// the shared fixture DB. Tests run sequentially within the class via xunit's
+/// per-class collection serialisation.
+/// </remarks>
+public class AuditLogPartitionMaintenanceTests : IClassFixture<MsSqlMigrationFixture>
+{
+    private readonly MsSqlMigrationFixture _fixture;
+
+    public AuditLogPartitionMaintenanceTests(MsSqlMigrationFixture fixture)
+    {
+        _fixture = fixture;
+    }
+
+    private ScadaLinkDbContext CreateContext() =>
+        new(new DbContextOptionsBuilder<ScadaLinkDbContext>()
+            .UseSqlServer(_fixture.ConnectionString).Options);
+
+    private AuditLogPartitionMaintenance NewMaintenance(ScadaLinkDbContext ctx) =>
+        new(ctx, NullLogger<AuditLogPartitionMaintenance>.Instance);
+
+    /// <summary>
+    /// Computes the lookahead-in-months required to fall strictly inside the
+    /// already-covered boundary range. Picks something well below the
+    /// distance from "now" to the current max — guaranteed not to need any
+    /// new SPLIT.
+    /// </summary>
+    private static int LookaheadInsideExistingRange(DateTime max)
+    {
+        var now = DateTime.UtcNow;
+        // (max - now) in whole months, minus a 1-month safety margin so we
+        // never accidentally hit the boundary horizon edge case.
+        var months = ((max.Year - now.Year) * 12) + max.Month - now.Month - 1;
+        return Math.Max(1, months);
+    }
+
+    /// <summary>
+    /// Computes the lookahead-in-months required to add exactly
+    /// <paramref name="extraBoundaries"/> new boundaries past the current max.
+    /// </summary>
+    /// <remarks>
+    /// EnsureLookaheadAsync defines horizon =
+    /// <c>NormalizeToFirstOfMonth(UtcNow) + lookaheadMonths</c>. The new
+    /// boundaries it issues are first-of-month values strictly greater than
+    /// max, up to and including horizon. So
+    /// <c>lookaheadMonths = monthsBetween(NormalizeToFirstOfMonth(UtcNow), max) + extraBoundaries</c>
+    /// is the exact value that lands horizon on <c>max + extraBoundaries</c>
+    /// months.
+    /// </remarks>
+    private static int LookaheadForExtraBoundaries(DateTime max, int extraBoundaries)
+    {
+        var nowFirstOfMonth = FirstOfNextMonth(DateTime.UtcNow);
+        var monthsToMax = ((max.Year - nowFirstOfMonth.Year) * 12) + max.Month - nowFirstOfMonth.Month;
+        return monthsToMax + extraBoundaries;
+    }
+
+    private static DateTime FirstOfNextMonth(DateTime instant)
+    {
+        var firstOfThisMonth = new DateTime(instant.Year, instant.Month, 1, 0, 0, 0, DateTimeKind.Utc);
+        return firstOfThisMonth.AddMonths(1);
+    }
+
+    [SkippableFact]
+    public async Task EnsureLookahead_AlreadyHasFutureRange_NoSplit_ReturnsEmpty()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var ctx = CreateContext();
+        var maintenance = NewMaintenance(ctx);
+
+        var max = await maintenance.GetMaxBoundaryAsync();
+        Assert.NotNull(max);
+
+        // Pick a lookahead small enough that horizon (NormalizeToFirstOfMonth(now)
+        // + lookahead) lands well INSIDE the already-covered range — no SPLIT
+        // should fire.
+        var lookahead = LookaheadInsideExistingRange(max.Value);
+
+        var added = await maintenance.EnsureLookaheadAsync(lookahead);
+
+        Assert.Empty(added);
+
+        // Sanity: the max boundary is unchanged after the no-op call.
+        var maxAfter = await maintenance.GetMaxBoundaryAsync();
+        Assert.Equal(max, maxAfter);
+    }
+
+    [SkippableFact]
+    public async Task EnsureLookahead_NeedsOneMoreBoundary_Splits_Returns1Boundary()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var ctx = CreateContext();
+        var maintenance = NewMaintenance(ctx);
+
+        var maxBefore = await maintenance.GetMaxBoundaryAsync();
+        Assert.NotNull(maxBefore);
+
+        var lookahead = LookaheadForExtraBoundaries(maxBefore.Value, extraBoundaries: 1);
+        var expectedAdded = maxBefore.Value.AddMonths(1);
+
+        var added = await maintenance.EnsureLookaheadAsync(lookahead);
+
+        Assert.Single(added);
+        Assert.Equal(expectedAdded, added[0]);
+
+        var maxAfter = await maintenance.GetMaxBoundaryAsync();
+        Assert.Equal(expectedAdded, maxAfter);
+    }
+
+    [SkippableFact]
+    public async Task EnsureLookahead_NeedsThreeBoundaries_Splits_Returns3Boundaries()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var ctx = CreateContext();
+        var maintenance = NewMaintenance(ctx);
+
+        var maxBefore = await maintenance.GetMaxBoundaryAsync();
+        Assert.NotNull(maxBefore);
+
+        var lookahead = LookaheadForExtraBoundaries(maxBefore.Value, extraBoundaries: 3);
+
+        var added = await maintenance.EnsureLookaheadAsync(lookahead);
+
+        Assert.Equal(3, added.Count);
+        Assert.Equal(maxBefore.Value.AddMonths(1), added[0]);
+        Assert.Equal(maxBefore.Value.AddMonths(2), added[1]);
+        Assert.Equal(maxBefore.Value.AddMonths(3), added[2]);
+
+        var maxAfter = await maintenance.GetMaxBoundaryAsync();
+        Assert.Equal(maxBefore.Value.AddMonths(3), maxAfter);
+    }
+
+    [SkippableFact]
+    public async Task EnsureLookahead_BoundaryAlreadyExists_NoError_Idempotent()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var ctx1 = CreateContext();
+        var m1 = NewMaintenance(ctx1);
+
+        var maxStart = await m1.GetMaxBoundaryAsync();
+        Assert.NotNull(maxStart);
+
+        // First call: add one boundary.
+        var lookahead = LookaheadForExtraBoundaries(maxStart.Value, extraBoundaries: 1);
+        var firstAdded = await m1.EnsureLookaheadAsync(lookahead);
+        Assert.Single(firstAdded);
+
+        // Second call: the boundary just added is now part of pf_AuditLog_Month,
+        // so the same lookahead value should be a no-op — no exception, no
+        // duplicate SPLIT.
+        await using var ctx2 = CreateContext();
+        var m2 = NewMaintenance(ctx2);
+        var secondAdded = await m2.EnsureLookaheadAsync(lookahead);
+
+        Assert.Empty(secondAdded);
+
+        // The max boundary is unchanged across the second call.
+        var maxAfter = await m2.GetMaxBoundaryAsync();
+        Assert.Equal(firstAdded[0], maxAfter);
+    }
+}

tests/ScadaLink.ConfigurationDatabase.Tests/Repositories/AuditLogRepositoryTests.cs

@@ -1,3 +1,4 @@
+using Microsoft.Data.SqlClient;
 using Microsoft.EntityFrameworkCore;
 using ScadaLink.Commons.Entities.Audit;
 using ScadaLink.Commons.Types.Audit;
@@ -309,21 +310,221 @@ public class AuditLogRepositoryTests : IClassFixture<MsSqlMigrationFixture>
         Assert.True(events.Select(e => e.EventId).ToHashSet().SetEquals(allIds));
     }
 
+    // ------------------------------------------------------------------------
+    // M6-T4 Bundle C: SwitchOutPartitionAsync drop-and-rebuild integration tests
+    // ------------------------------------------------------------------------
+    //
+    // The partition-switch path replaces M1's NotSupportedException stub with
+    // the production drop-DROP-INDEX → CREATE-staging → SWITCH PARTITION →
+    // DROP-staging → CREATE-INDEX dance documented in alog.md §4. These tests
+    // verify the side effects an outsider can observe:
+    //   * rows in the targeted month are removed
+    //   * rows in OTHER months are NOT touched
+    //   * UX_AuditLog_EventId still exists after a successful switch
+    //   * InsertIfNotExistsAsync's first-write-wins idempotency still holds
+    //     after a switch (the rebuilt index is real)
+    //   * a thrown SqlException leaves UX_AuditLog_EventId rebuilt (the CATCH
+    //     branch's recovery path runs)
+
     [SkippableFact]
-    public async Task SwitchOutPartitionAsync_ThrowsNotSupported_ForM1()
+    public async Task SwitchOutPartitionAsync_OldPartition_RemovesRows_NewPartitionsKept()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = NewSiteId();
+        await using var context = CreateContext();
+        var repo = new AuditLogRepository(context);
+
+        // Three distinct months — Jan, Feb, Mar 2026 — so the switch on Jan's
+        // boundary purges exactly one month's worth of rows. Boundary values
+        // come from the partition function's pre-seeded list (alog.md §4).
+        var janEvt = NewEvent(siteId, occurredAtUtc: new DateTime(2026, 1, 15, 10, 0, 0, DateTimeKind.Utc));
+        var febEvt = NewEvent(siteId, occurredAtUtc: new DateTime(2026, 2, 15, 10, 0, 0, DateTimeKind.Utc));
+        var marEvt = NewEvent(siteId, occurredAtUtc: new DateTime(2026, 3, 15, 10, 0, 0, DateTimeKind.Utc));
+        await repo.InsertIfNotExistsAsync(janEvt);
+        await repo.InsertIfNotExistsAsync(febEvt);
+        await repo.InsertIfNotExistsAsync(marEvt);
+
+        // Boundary value '2026-01-01' identifies the January 2026 partition under
+        // RANGE RIGHT semantics ($PARTITION returns the partition into which the
+        // boundary value itself falls — the partition whose lower bound is the
+        // boundary).
+        await repo.SwitchOutPartitionAsync(new DateTime(2026, 1, 1, 0, 0, 0, DateTimeKind.Utc));
+
+        await using var readContext = CreateContext();
+        var remaining = await readContext.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == siteId)
+            .ToListAsync();
+
+        Assert.DoesNotContain(remaining, e => e.EventId == janEvt.EventId);
+        Assert.Contains(remaining, e => e.EventId == febEvt.EventId);
+        Assert.Contains(remaining, e => e.EventId == marEvt.EventId);
+    }
+
+    [SkippableFact]
+    public async Task SwitchOutPartitionAsync_RebuildsUxIndex_AfterSwitch()
     {
         Skip.IfNot(_fixture.Available, _fixture.SkipReason);
 
         await using var context = CreateContext();
         var repo = new AuditLogRepository(context);
 
-        // The partition-switch path is intentionally blocked in M1 because
-        // UX_AuditLog_EventId is non-aligned. The drop-and-rebuild dance ships
-        // with the M6 purge actor.
-        var ex = await Assert.ThrowsAsync<NotSupportedException>(
-            () => repo.SwitchOutPartitionAsync(new DateTime(2026, 2, 1, 0, 0, 0, DateTimeKind.Utc)));
+        // Pick a different month per test so successive test runs (which share
+        // the fixture's MSSQL database) don't tread on each other.
+        await repo.SwitchOutPartitionAsync(new DateTime(2026, 4, 1, 0, 0, 0, DateTimeKind.Utc));
 
-        Assert.Contains("M6", ex.Message, StringComparison.OrdinalIgnoreCase);
+        await using var verifyContext = CreateContext();
+        var indexExists = await ScalarAsync<int>(
+            verifyContext,
+            "SELECT COUNT(*) FROM sys.indexes " +
+            "WHERE name = 'UX_AuditLog_EventId' AND object_id = OBJECT_ID('dbo.AuditLog');");
+        Assert.Equal(1, indexExists);
+    }
+
+    [SkippableFact]
+    public async Task SwitchOutPartitionAsync_InsertIfNotExistsAsync_StillEnforcesFirstWriteWins_AfterSwitch()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = NewSiteId();
+        await using var context = CreateContext();
+        var repo = new AuditLogRepository(context);
+
+        // Pre-existing row in May 2026 — must survive a switch on a different
+        // (older) partition.
+        var preExisting = NewEvent(siteId, occurredAtUtc: new DateTime(2026, 5, 20, 9, 0, 0, DateTimeKind.Utc));
+        await repo.InsertIfNotExistsAsync(preExisting);
+
+        // Switch out the June 2026 partition (different month, empty).
+        await repo.SwitchOutPartitionAsync(new DateTime(2026, 6, 1, 0, 0, 0, DateTimeKind.Utc));
+
+        // Re-attempting the same EventId after the switch must STILL be a no-op
+        // (UX_AuditLog_EventId is the index that enables idempotency; if the
+        // rebuild left it broken, this insert would silently produce a duplicate
+        // row and the count assertion below would catch it).
+        var dup = preExisting with { ErrorMessage = "second-should-be-ignored-after-switch" };
+        await repo.InsertIfNotExistsAsync(dup);
+
+        await using var readContext = CreateContext();
+        var rows = await readContext.Set<AuditEvent>()
+            .Where(e => e.SourceSiteId == siteId)
+            .ToListAsync();
+
+        Assert.Single(rows);
+        Assert.Equal(preExisting.EventId, rows[0].EventId);
+        // First-write-wins: the original ErrorMessage (null) survives.
+        Assert.Null(rows[0].ErrorMessage);
+    }
+
+    [SkippableFact]
+    public async Task SwitchOutPartitionAsync_PartialFailure_RebuildsUxIndex_RaisesException()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        await using var context = CreateContext();
+        var repo = new AuditLogRepository(context);
+
+        // Force a deterministic switch failure with an inbound FOREIGN KEY:
+        // ALTER TABLE … SWITCH refuses to move rows out of a partition that's
+        // referenced by an FK from another table, raising msg 4928
+        // ("ALTER TABLE SWITCH statement failed because target table … has a
+        // foreign key …"). The CATCH branch then rolls back and rebuilds the
+        // unique index — which the assertion below verifies.
+        //
+        // The probe table is uniquely named with a guid suffix so reruns of
+        // this test inside the same fixture DB never collide. We clean it up
+        // in the finally so the constraint never leaks into other tests.
+        var probeTable = $"AuditFkProbe_{Guid.NewGuid():N}".Substring(0, 32);
+        await using (var setup = new SqlConnection(_fixture.ConnectionString))
+        {
+            await setup.OpenAsync();
+            await using var cmd = setup.CreateCommand();
+            // Composite FK references AuditLog's composite PK (EventId, OccurredAtUtc).
+            cmd.CommandText =
+                $"CREATE TABLE dbo.[{probeTable}] ( " +
+                $"  EventId uniqueidentifier NOT NULL, " +
+                $"  OccurredAtUtc datetime2(7) NOT NULL, " +
+                $"  CONSTRAINT FK_{probeTable}_AuditLog FOREIGN KEY (EventId, OccurredAtUtc) " +
+                $"    REFERENCES dbo.AuditLog(EventId, OccurredAtUtc));";
+            await cmd.ExecuteNonQueryAsync();
+        }
+
+        try
+        {
+            var ex = await Assert.ThrowsAnyAsync<SqlException>(
+                () => repo.SwitchOutPartitionAsync(new DateTime(2026, 9, 1, 0, 0, 0, DateTimeKind.Utc)));
+            // Smoke-check the message references the SWITCH statement so we
+            // know we hit the engineered failure, not some unrelated error.
+            Assert.Contains("SWITCH", ex.Message, StringComparison.OrdinalIgnoreCase);
+        }
+        finally
+        {
+            // Always drop the probe table so the FK is gone before the next
+            // test runs against the shared fixture.
+            await using var cleanup = new SqlConnection(_fixture.ConnectionString);
+            await cleanup.OpenAsync();
+            await using var cmd = cleanup.CreateCommand();
+            cmd.CommandText =
+                $"IF OBJECT_ID('dbo.[{probeTable}]', 'U') IS NOT NULL DROP TABLE dbo.[{probeTable}];";
+            await cmd.ExecuteNonQueryAsync();
+        }
+
+        // The CATCH block in the production SQL guarantees UX_AuditLog_EventId
+        // is rebuilt regardless of which step failed inside the TRY.
+        await using var verifyContext = CreateContext();
+        var indexExists = await ScalarAsync<int>(
+            verifyContext,
+            "SELECT COUNT(*) FROM sys.indexes " +
+            "WHERE name = 'UX_AuditLog_EventId' AND object_id = OBJECT_ID('dbo.AuditLog');");
+        Assert.Equal(1, indexExists);
+    }
+
+    // ------------------------------------------------------------------------
+    // M6-T4 Bundle C: GetPartitionBoundariesOlderThanAsync
+    // ------------------------------------------------------------------------
+
+    [SkippableFact]
+    public async Task GetPartitionBoundariesOlderThanAsync_ReturnsBoundaries_WithMaxOccurredOlderThanThreshold()
+    {
+        Skip.IfNot(_fixture.Available, _fixture.SkipReason);
+
+        var siteId = NewSiteId();
+        await using var context = CreateContext();
+        var repo = new AuditLogRepository(context);
+
+        // Seed events in two months: July 2026 (old) and August 2026 (new).
+        await repo.InsertIfNotExistsAsync(NewEvent(siteId, occurredAtUtc: new DateTime(2026, 7, 10, 0, 0, 0, DateTimeKind.Utc)));
+        await repo.InsertIfNotExistsAsync(NewEvent(siteId, occurredAtUtc: new DateTime(2026, 8, 10, 0, 0, 0, DateTimeKind.Utc)));
+
+        // Threshold = Aug 1 2026 — July partition's MAX (July 10) is older;
+        // August partition's MAX (August 10) is newer. We expect only the July
+        // boundary back.
+        var threshold = new DateTime(2026, 8, 1, 0, 0, 0, DateTimeKind.Utc);
+        var boundaries = await repo.GetPartitionBoundariesOlderThanAsync(threshold);
+
+        // The repo may also return EARLIER boundaries that have no data (their
+        // MAX is NULL → treated as "no data, nothing to purge" by the contract).
+        // We only assert the inclusion/exclusion that matters for our seeded
+        // rows.
+        Assert.Contains(new DateTime(2026, 7, 1, 0, 0, 0, DateTimeKind.Utc), boundaries);
+        Assert.DoesNotContain(new DateTime(2026, 8, 1, 0, 0, 0, DateTimeKind.Utc), boundaries);
+    }
+
+    private async Task<T> ScalarAsync<T>(ScadaLinkDbContext context, string sql)
+    {
+        var conn = context.Database.GetDbConnection();
+        if (conn.State != System.Data.ConnectionState.Open)
+        {
+            await conn.OpenAsync();
+        }
+        await using var cmd = conn.CreateCommand();
+        cmd.CommandText = sql;
+        var result = await cmd.ExecuteScalarAsync();
+        if (result is null || result is DBNull)
+        {
+            return default!;
+        }
+        return (T)Convert.ChangeType(result, typeof(T) == typeof(string) ? typeof(string) : Nullable.GetUnderlyingType(typeof(T)) ?? typeof(T))!;
     }
 
     // --- helpers ------------------------------------------------------------

tests/ScadaLink.HealthMonitoring.Tests/SiteAuditBacklogMetricTests.cs

@@ -0,0 +1,73 @@
+using ScadaLink.Commons.Types;
+
+namespace ScadaLink.HealthMonitoring.Tests;
+
+/// <summary>
+/// Bundle E (M6-T6) regression coverage. The site-side audit-log SQLite writer
+/// exposes a backlog snapshot (<c>SiteAuditBacklogSnapshot</c>) via the
+/// <c>ISiteAuditQueue.GetBacklogStatsAsync</c> surface. A periodic
+/// <c>SiteAuditBacklogReporter</c> hosted service polls that snapshot and
+/// pushes it into the collector via <see cref="ISiteHealthCollector.UpdateSiteAuditBacklog"/>
+/// so the next <see cref="ISiteHealthCollector.CollectReport"/> includes it in
+/// the report payload as <c>SiteAuditBacklog</c>. Unlike the
+/// SiteAuditWriteFailures / AuditRedactionFailure interval counters, the
+/// backlog snapshot is not reset on collect — the field carries forward
+/// whatever the most recent refresh pushed in.
+/// </summary>
+public class SiteAuditBacklogMetricTests
+{
+    private readonly SiteHealthCollector _collector = new();
+
+    [Fact]
+    public void Update_Then_CollectReport_IncludesBacklog()
+    {
+        var snapshot = new SiteAuditBacklogSnapshot(
+            PendingCount: 42,
+            OldestPendingUtc: new DateTime(2026, 5, 20, 10, 0, 0, DateTimeKind.Utc),
+            OnDiskBytes: 1234567);
+
+        _collector.UpdateSiteAuditBacklog(snapshot);
+
+        var report = _collector.CollectReport("site-1");
+
+        Assert.Equal(snapshot, report.SiteAuditBacklog);
+    }
+
+    [Fact]
+    public void Report_Payload_Includes_SiteAuditBacklog_AsNullByDefault()
+    {
+        // No refresh has been pushed yet — the report carries null so the
+        // central UI can distinguish "no data yet" from "queue empty".
+        var report = _collector.CollectReport("site-1");
+
+        Assert.Null(report.SiteAuditBacklog);
+    }
+
+    [Fact]
+    public void CollectReport_DoesNotReset_SiteAuditBacklog()
+    {
+        // Backlog snapshot is a point-in-time reading, not a per-interval
+        // counter — successive CollectReport calls before the next
+        // SiteAuditBacklogReporter tick MUST keep returning the same snapshot
+        // so a slow refresh cadence doesn't blank the central dashboard.
+        var snapshot = new SiteAuditBacklogSnapshot(
+            PendingCount: 7,
+            OldestPendingUtc: null,
+            OnDiskBytes: 8192);
+
+        _collector.UpdateSiteAuditBacklog(snapshot);
+
+        var first = _collector.CollectReport("site-1");
+        var second = _collector.CollectReport("site-1");
+
+        Assert.Equal(snapshot, first.SiteAuditBacklog);
+        Assert.Equal(snapshot, second.SiteAuditBacklog);
+    }
+
+    [Fact]
+    public void Update_With_Null_Throws_ArgumentNullException()
+    {
+        Assert.Throws<ArgumentNullException>(
+            () => _collector.UpdateSiteAuditBacklog(null!));
+    }
+}

tests/ScadaLink.SiteRuntime.Tests/Actors/DeploymentManagerRedeployTests.cs

@@ -71,6 +71,7 @@ public class DeploymentManagerRedeployTests : TestKit, IDisposable
         public void IncrementDeadLetter() { }
         public void IncrementSiteAuditWriteFailures() { }
         public void IncrementAuditRedactionFailure() { }
+        public void UpdateSiteAuditBacklog(ScadaLink.Commons.Types.SiteAuditBacklogSnapshot snapshot) { }
         public void UpdateConnectionHealth(string connectionName, ConnectionHealth health) { }
         public void RemoveConnection(string connectionName) { }
         public void UpdateTagResolution(string connectionName, int totalSubscribed, int successfullyResolved) { }