The alert engine takes per-server and per-database threshold overrides.

Darling/Darling.Tests/AlertEngineTests.cs

@@ -459,6 +459,8 @@ into the process-wide AlertReadFailureCounter.Shared. */
         public AlertReadFailureCounter? ReadFailures { get; set; }
         /// <summary>#5715 part B: the alert-history reader that seeds the escalation tier; null (the default) leaves the engine memory-only.</summary>
         public Func<EscalationSeedQuery, CancellationToken, Task<AlertSeverityLevel?>>? EscalationSeedReader { get; set; }
+        /// <summary>#5910: the threshold overrides; null (the default) is the engine before #5910, so every pin above is unchanged.</summary>
+        public IAlertThresholdOverrides? Overrides { get; set; }
 
         /* #3497: the Agent-job resolver seam, null by default so every pre-existing pin builds the
            pre-#3497 engine. Tests set AgentJobNames (the answer map), flip AgentJobResolverThrows for
@@ -489,7 +491,8 @@ into the process-wide AlertReadFailureCounter.Shared. */
             logger: null,
             utcNow: () => Now,
             readFailures: ReadFailures,
-            escalationSeedReader: EscalationSeedReader);
+            escalationSeedReader: EscalationSeedReader,
+            overrides: Overrides);
 
         /* #3282: every snapshot gets a DISTINCT, increasing CPU sample instant unless the test says
            otherwise, because that is the realistic case — the sweep sees a new ring-buffer sample each
@@ -3885,7 +3888,7 @@ public async Task LowDisk_GradesCriticallyLowBreaches_AndAStandingBreachDoesNotR
 
     /* ---------------- database file growth (#2349; #3539 A8c: the rise is MB per HOUR) ---------------- */
 
-    private static DatabaseFileGrowthInfo GrowingFile(double growthMb, double windowMinutes) => new()
+    internal static DatabaseFileGrowthInfo GrowingFile(double growthMb, double windowMinutes) => new()
     {
         DatabaseName = "tempdb", FileName = "tempdev", PhysicalName = @"D:\data\tempdev.mdf", FileTypeDesc = "ROWS",
         TotalSizeMb = 90_000, GrowthMb = growthMb, GrowthWindowMinutes = windowMinutes,
@@ -3967,12 +3970,12 @@ public async Task FileGrowth_ARateUnderTheBar_IsSilentOnAnyLookback(int lookback
     /// <summary>Two consecutive hourly collections of <c>database_size_stats</c>, so the pins below replay the
     /// shape the issue describes rather than an invented one: a rise observed at the top of the hour, twelve
     /// five-minute cooldowns of re-reads against the same two rows, the next collection an hour later.</summary>
-    private static readonly DateTime HourlyCollection0 = new(2026, 9, 18, 6, 0, 0, DateTimeKind.Utc);
-    private static readonly DateTime HourlyCollection1 = HourlyCollection0.AddHours(1);
+    internal static readonly DateTime HourlyCollection0 = new(2026, 9, 18, 6, 0, 0, DateTimeKind.Utc);
+    internal static readonly DateTime HourlyCollection1 = HourlyCollection0.AddHours(1);
 
     /// <summary>A rise-only file (2% of a 4 TB volume — the level gate cannot see it) stamped with the collection
     /// that produced it. 20 GB in the 60-minute window is twice the default 10,240 MB/hr bar.</summary>
-    private static DatabaseFileGrowthInfo RiseOnlyFile(DateTime? observedAt, double growthMb = 20_480, string fileName = "tempdev")
+    internal static DatabaseFileGrowthInfo RiseOnlyFile(DateTime? observedAt, double growthMb = 20_480, string fileName = "tempdev")
     {
         var f = GrowingFile(growthMb, windowMinutes: 60);
         f.FileName = fileName;
@@ -3982,7 +3985,7 @@ private static DatabaseFileGrowthInfo RiseOnlyFile(DateTime? observedAt, double
 
     /// <summary>A level-only file: 80% of a small volume and not growing at all. The standing-level shape that
     /// re-fires on the cooldown by design.</summary>
-    private static DatabaseFileGrowthInfo LevelOnlyFile(DateTime? observedAt) => new()
+    internal static DatabaseFileGrowthInfo LevelOnlyFile(DateTime? observedAt) => new()
     {
         DatabaseName = "Sales", FileName = "Sales_log", PhysicalName = @"L:\log\Sales_log.ldf", FileTypeDesc = "LOG",
         TotalSizeMb = 400_000, GrowthMb = 0, GrowthWindowMinutes = 60,

Darling/Darling.Tests/AlertReadFailureSurfaceTests.cs

@@ -915,7 +915,9 @@ real failure rather than a matcher that never matches anything. */
     /// </summary>
     private static readonly (string Path, int Counted, int Exempt)[] s_wholeFileScopes =
     {
-        (Path.Combine("PerformanceMonitor.Alerting", "AlertEngine.cs"), 18, 6),
+        (Path.Combine("PerformanceMonitor.Alerting", "AlertEngine.cs"), 18, 7),
+        /* 7th AlertEngine exempt since #5910: the threshold-override fallback in Threshold(), an in-memory lookup and
+           not a read; the check it serves still reads, and counts, exactly as before. */
         /* 14th exempt since #4215: EvaluateStoreSettingsAsync's wrapper catch — the store-settings
            self-alert's report is a parameter, exactly like its StaleMute/WebTls siblings. 15th exempt
            since #4299: EvaluateRawPurgeOverHorizonAsync's wrapper catch — the raw-purge-over-horizon
@@ -1087,6 +1089,7 @@ holds the standing alert on a failed read. */
         ["delivery stamp could not be written"] = "a write (#3580): the daily document was already delivered and process memory already gates it; the dropped stamp costs one re-announcement at the next restart and never a delivery - the stamp READ beside it is the read, and it is counted",
         ["Store statement history is still failing"] = "the Debug repeat of a telemetry write sweep (#5097); the first failure warns, no alert is judged on its result",
         ["Collection log hourly rollup failed"] = "a maintenance write (#5810): the pass upserts the hourly rollup and no alert is judged on its result; a failed pass is redone by the next, and get_collection_health reads the hours after the newest rolled hour from the raw log",
+        ["The threshold override for"] = "an in-memory override lookup (#5910), not a read: the global value applies and the check's own read is counted as before",
         ["Read-latency flush failed"] ="the read-latency histogram flush writes instrumentation; losing an hour's rows can't hide an alert condition",
     };
 
@@ -1280,8 +1283,9 @@ WebTlsCertificateState report the web host publishes - there is no store read to
            wrapper catch, handed its report as a parameter. 39th since #5810: the collection log hourly rollup's catch, a
            maintenance write no alert is judged on, redone by the next pass. 40th since #5867: the Compose Rollup Hours Stuck
            self-alert's wrapper catch, handed its failure record as a parameter. 41st since #5867: the worker's read of that
-           record, which holds the standing alert on a failure. */
-        Assert.Equal(41, totalExempt);
+           record, which holds the standing alert on a failure. 42nd since #5910: the engine's threshold-override
+           fallback, an in-memory lookup and not a read. */
+        Assert.Equal(42, totalExempt);
 
         /* Every exemption in the table is actually used. An exemption for a message that no longer exists
            is a hole this pin would otherwise keep open indefinitely — the shape that lets a real new catch