Fix false-positive slow alerts: require 3 consecutive checks + use P95 baseline

Previously a single check above 2× average response time fired an alert.
This caused false positives from normal network variance.

- scheduler.js: track consecutive slow checks per resource; only alert after
  3 in a row (resets to 0 when response time normalises)
- monitorService.js: calculateResponseTimeBaseline now returns P95 instead
  of mean, making 2× baseline a much more realistic slow threshold

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

server/monitorService.js

@@ -765,9 +765,10 @@ class MonitorService {
       return null;
     }
 
-    // Return average of response times
-    const sum = checks.reduce((acc, c) => acc + (c.response_time || 0), 0);
-    return Math.round(sum / checks.length);
+    // Use P95 so the baseline reflects realistic worst-case, not the average.
+    // Threshold is 2× P95, making it very unlikely to trigger on normal variance.
+    const p95Index = Math.min(Math.floor(checks.length * 0.95), checks.length - 1);
+    return Math.round(checks[p95Index].response_time);
   }
 
   // Check if response time is anomalous