fix(backend): agent-silence sweep must compare clock frames it owns

LastEventAt is the maximum AGENT-supplied event timestamp and therefore lives in
the device clock frame. The silence sweep compared it against a server-derived
cutoff, mixing two frames.

Measured over 11,068 sessions: 26 carry a genuine CMTrace timezone skew on their
IME-derived events, from -17 h to +1 h (a further 27 carry plain device-clock
offsets). At -17 h a live agent looks silent the moment it starts and is marked
Stalled with a customer-visible "Agent silent for 1020min"; at +1 h the sweep is
blind for an hour.

Adds SessionSummary.LastIngestAt, stamped from the server clock in
IncrementSessionEventCountAsync — reaching that method IS contact from the agent,
so it is written unconditionally rather than gated on latestEventTimestamp.

The sweep's server-side filter now spans both frames (LastIngestAt lt X or
LastEventAt lt X): filtering on LastIngestAt alone would silently drop every row
written before the field existed, because Table Storage does not match a filter
against a missing property. The authoritative decision moves into
ServerFrameLastContact() (LastIngestAt ?? LastEventAt ?? StartedAt) and is applied
in both stages — before the Stalled marker and in the stage-2 alive guard.

Deliberately not mirrored into SessionsIndex: the sweep reads Sessions, and the
index manifest is pinned bidirectionally by tests.

shared-manifests.json + .generated.ts regenerated — SharedManifestParityTests is
the drift brake and correctly went red on the new field.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

src/Backend/AutopilotMonitor.Functions.Tests/MaintenanceServerFrameLastContactTests.cs

@@ -0,0 +1,75 @@
+using System;
+using AutopilotMonitor.Functions.Services;
+using AutopilotMonitor.Shared.Models;
+using Xunit;
+
+namespace AutopilotMonitor.Functions.Tests
+{
+    /// <summary>
+    /// Pins the clock-frame contract of the agent-silence sweep.
+    ///
+    /// <para>
+    /// <c>LastEventAt</c> is the maximum AGENT-supplied event timestamp and therefore lives in the
+    /// device clock frame — it carries the device's clock error and, for CMTrace-derived events,
+    /// the writing process's timezone belief. Field measurement over 11k sessions found skews from
+    /// -17 h to +1 h. Comparing that against a server-derived silence cutoff marked live agents
+    /// Stalled ("Agent silent for 1020min") and left genuinely silent ones undetected for the
+    /// duration of the skew. <c>LastIngestAt</c> is stamped from the server clock on every ingest
+    /// and is the only value such a comparison may use.
+    /// </para>
+    /// </summary>
+    public class MaintenanceServerFrameLastContactTests
+    {
+        private static readonly DateTime Now = new(2026, 08, 20, 12, 00, 00, DateTimeKind.Utc);
+
+        private static SessionSummary Session(DateTime? lastIngestAt, DateTime? lastEventAt, DateTime startedAt)
+            => new() { LastIngestAt = lastIngestAt, LastEventAt = lastEventAt, StartedAt = startedAt };
+
+        [Fact]
+        public void PrefersServerFrameIngestStamp_OverDeviceFrameEventStamp()
+        {
+            var s = Session(lastIngestAt: Now.AddMinutes(-1), lastEventAt: Now.AddHours(-17), startedAt: Now.AddHours(-2));
+            Assert.Equal(Now.AddMinutes(-1), MaintenanceService.ServerFrameLastContact(s));
+        }
+
+        [Fact]
+        public void NegativeSkew_DoesNotLookSilent_WhenIngestIsFresh()
+        {
+            // The exact field shape: IME-derived events landed 17 h in the past, so LastEventAt is
+            // far behind the 2 h silence cutoff while the agent is demonstrably still reporting.
+            var silenceCutoff = Now.AddHours(-2);
+            var s = Session(lastIngestAt: Now.AddSeconds(-30), lastEventAt: Now.AddHours(-17), startedAt: Now.AddHours(-1));
+
+            Assert.True(MaintenanceService.ServerFrameLastContact(s) > silenceCutoff,
+                "a session whose ingest stamp is 30 s old must never be classified as agent-silent");
+        }
+
+        [Fact]
+        public void PositiveSkew_DoesNotMaskSilence_WhenIngestIsStale()
+        {
+            // Mirror case: a +1 h skew pushes LastEventAt into the future, which previously made the
+            // sweep blind for an hour. The server frame still sees the silence.
+            var silenceCutoff = Now.AddHours(-2);
+            var s = Session(lastIngestAt: Now.AddHours(-3), lastEventAt: Now.AddHours(1), startedAt: Now.AddHours(-4));
+
+            Assert.True(MaintenanceService.ServerFrameLastContact(s) < silenceCutoff,
+                "a session with no ingest for 3 h must be detected as silent regardless of a future-dated event stamp");
+        }
+
+        [Fact]
+        public void FallsBackToEventStamp_OnRowsPredatingTheIngestStamp()
+        {
+            // Rollout safety: rows written before LastIngestAt existed must keep the old behaviour
+            // rather than dropping out of the sweep entirely.
+            var s = Session(lastIngestAt: null, lastEventAt: Now.AddHours(-3), startedAt: Now.AddHours(-5));
+            Assert.Equal(Now.AddHours(-3), MaintenanceService.ServerFrameLastContact(s));
+        }
+
+        [Fact]
+        public void FallsBackToStartedAt_WhenNoEventStampExists()
+        {
+            var s = Session(lastIngestAt: null, lastEventAt: null, startedAt: Now.AddHours(-4));
+            Assert.Equal(Now.AddHours(-4), MaintenanceService.ServerFrameLastContact(s));
+        }
+    }
+}

src/Backend/AutopilotMonitor.Functions/Services/MaintenanceService.cs

@@ -351,8 +351,16 @@ namespace AutopilotMonitor.Functions.Services
                         int silentMarked = 0;
                         foreach (var silent in silentSessions)
                         {
-                            var lastEventAt = silent.LastEventAt ?? silent.StartedAt;
-                            var silentMinutes = (int)(now - lastEventAt).TotalMinutes;
+                            // The query is a permissive two-frame pre-filter; the decision is made
+                            // here on the SERVER frame. LastEventAt carries the device clock and,
+                            // for IME-derived events, a timezone skew (field-measured: -17 h to
+                            // +1 h). Trusting it marks a live agent Stalled — and puts a bogus
+                            // "silent for 1020min" in front of the customer.
+                            var lastContactAt = ServerFrameLastContact(silent);
+                            if (lastContactAt > silenceCutoff)
+                                continue; // provably alive on the server clock
+
+                            var silentMinutes = (int)(now - lastContactAt).TotalMinutes;
                             await _sessionRepo.UpdateSessionStatusAsync(
                                 silent.TenantId,
                                 silent.SessionId,
@@ -409,7 +417,7 @@ namespace AutopilotMonitor.Functions.Services
                             // legitimately exceed SessionTimeoutHours) and must never be terminalized here
                             // — without this, an actively-installing 6h enrollment could be classified
                             // Incomplete mid-run purely because it outlived the 5h window.
-                            if (session.LastEventAt.HasValue && session.LastEventAt.Value > silenceCutoff)
+                            if (ServerFrameLastContact(session) > silenceCutoff)
                                 continue;
 
                             // Fast path: an AwaitingUser session still inside the grace window needs no
@@ -774,5 +782,26 @@ namespace AutopilotMonitor.Functions.Services
         /// Aggregates metrics for any missed days in the last 7 days, plus yesterday.
         /// Checks the UsageMetrics table for existing snapshots to avoid re-aggregation.
         /// </summary>
+
+        /// <summary>
+        /// Last contact with the agent, expressed in the SERVER clock frame — the only value a
+        /// server-derived cutoff may legitimately be compared against.
+        /// <para>
+        /// Preference order:
+        /// <list type="number">
+        /// <item><c>LastIngestAt</c> — stamped from the server clock on every ingest. Authoritative.</item>
+        /// <item><c>LastEventAt</c> — device frame; used only for rows written before
+        /// <c>LastIngestAt</c> existed, so the rollout has no blind window. Carries the device's
+        /// clock error and any CMTrace timezone skew.</item>
+        /// <item><c>StartedAt</c> — last resort when a session has no event stamp at all.</item>
+        /// </list>
+        /// </para>
+        /// Mixing the two frames is what let a live agent be marked Stalled: a session whose
+        /// IME-derived events were skewed -17 h looked silent for 17 hours the moment it started.
+        /// </summary>
+        /// <remarks>internal (not private) purely as a unit-test seam — it is a pure function
+        /// over its argument, pinned by <c>MaintenanceServerFrameLastContactTests</c>.</remarks>
+        internal static DateTime ServerFrameLastContact(SessionSummary session)
+            => session.LastIngestAt ?? session.LastEventAt ?? session.StartedAt;
     }
 }

src/Backend/AutopilotMonitor.Functions/Services/TableStorageService.Maintenance.cs

@@ -698,9 +698,19 @@ namespace AutopilotMonitor.Functions.Services
 
                 var silenceCutoffStr = silenceCutoff.ToString("yyyy-MM-ddTHH:mm:ss");
                 var hardCutoffStr = hardCutoff.ToString("yyyy-MM-ddTHH:mm:ss");
+
+                // Candidate pre-filter across BOTH clock frames, deliberately permissive:
+                // LastIngestAt is the server frame (authoritative, but absent on rows written
+                // before it existed), LastEventAt the device frame (always present, but carries
+                // the device's clock error and any CMTrace timezone skew). An OR keeps a session
+                // in the candidate set when EITHER frame looks silent; the caller then makes the
+                // authoritative call on the server frame. Filtering on LastIngestAt alone would
+                // silently drop every pre-rollout row, since Table Storage does not match a
+                // filter against a missing property.
                 var filter = $"PartitionKey eq '{tenantId}' " +
                              $"and Status eq 'InProgress' " +
-                             $"and LastEventAt lt datetime'{silenceCutoffStr}Z' " +
+                             $"and (LastIngestAt lt datetime'{silenceCutoffStr}Z' " +
+                             $"or LastEventAt lt datetime'{silenceCutoffStr}Z') " +
                              $"and (StartedAt ge datetime'{hardCutoffStr}Z' or ResumedAt ge datetime'{hardCutoffStr}Z')";
 
                 var query = tableClient.QueryAsync<TableEntity>(filter: filter);
@@ -772,7 +782,7 @@ namespace AutopilotMonitor.Functions.Services
                 var select = new[]
                 {
                     "PartitionKey", "RowKey", "Status", "StartedAt", "ResumedAt", "LastEventAt",
-                    "SerialNumber", "IsPreProvisioned"
+                    "LastIngestAt", "SerialNumber", "IsPreProvisioned"
                 };
 
                 var sessions = new List<SessionSummary>();

src/Backend/AutopilotMonitor.Functions/Services/TableStorageService.Sessions.cs

@@ -2338,6 +2338,14 @@ namespace AutopilotMonitor.Functions.Services
                             update["LastEventAt"] = EnsureUtc(latestEventTimestamp.Value);
                     }
 
+                    // Server-frame ingest stamp. Reaching this method means a batch was just
+                    // received, so "now" on the SERVER clock is the honest answer to "when did we
+                    // last hear from this agent". LastEventAt above cannot answer that — it carries
+                    // the device's clock frame and its timezone skew. Written unconditionally (not
+                    // guarded by latestEventTimestamp) because an empty-timestamp batch is still
+                    // contact from the agent.
+                    update["LastIngestAt"] = EnsureUtc(DateTime.UtcNow);
+
                     // Increment script execution counters
                     if (platformScriptIncrement > 0)
                     {
@@ -3147,6 +3155,7 @@ namespace AutopilotMonitor.Functions.Services
                 // null as "CustomerSas" for back-compat.
                 DiagnosticsBlobDestination = entity.GetString("DiagnosticsBlobDestination"),
                 LastEventAt = SafeGetDateTime(entity, "LastEventAt"),
+                LastIngestAt = SafeGetDateTime(entity, "LastIngestAt"),
                 IsPreProvisioned = entity.GetBoolean("IsPreProvisioned") ?? false,
                 IsHybridJoin = entity.GetBoolean("IsHybridJoin") ?? false,
                 IsSelfDeployingProfile = entity.GetBoolean("IsSelfDeployingProfile") ?? false,

src/Shared/AutopilotMonitor.Shared/Models/SessionApiModels.cs

@@ -349,6 +349,21 @@ namespace AutopilotMonitor.Shared.Models
         /// </summary>
         public DateTime? LastEventAt { get; set; }
 
+        /// <summary>
+        /// SERVER-side time of the most recent event ingest for this session.
+        /// <para>
+        /// <see cref="LastEventAt"/> lives in the DEVICE clock frame (it is the maximum
+        /// agent-supplied event timestamp). Comparing it against a server-derived cutoff mixes
+        /// two clock frames: a device whose clock runs slow — or one whose IME-derived events
+        /// carry a timezone skew — is judged silent far too early or not at all (measured in the
+        /// field: skews from -17 h to +1 h, see tasks/todo.md). This field is stamped from the
+        /// server clock on every ingest, so silence detection can compare like with like.
+        /// </para>
+        /// Null for sessions that predate this field — callers MUST fall back to
+        /// <see cref="LastEventAt"/> so rollout has no blind window.
+        /// </summary>
+        public DateTime? LastIngestAt { get; set; }
+
         /// <summary>
         /// Whether this session used WhiteGlove (Pre-Provisioning).
         /// Set when a whiteglove_complete event is processed.

src/Web/autopilot-monitor-web/utils/shared-manifests.generated.ts

@@ -264,6 +264,10 @@ export const SHARED_MANIFEST = {
         "name": "lastEventAt",
         "optional": true
       },
+      {
+        "name": "lastIngestAt",
+        "optional": true
+      },
       {
         "name": "isPreProvisioned",
         "optional": false

src/Web/autopilot-monitor-web/utils/shared-manifests.json

@@ -260,6 +260,10 @@
         "name": "lastEventAt",
         "optional": true
       },
+      {
+        "name": "lastIngestAt",
+        "optional": true
+      },
       {
         "name": "isPreProvisioned",
         "optional": false