fix: exclude seed data from anomaly detection history when processing live metrics

When a seed scenario TTL expires, the ring buffer may still contain seed spike
data. If the agent loop then processes live metrics (no seedTtlExpiry), the old
spike history was causing false anomaly detection (e.g. txPool↑ medium when
live tx=1, because spike history caused monotonic/Z-score triggers).

Fix: filter history to exclude seed points (seedTtlExpiry set) when the current
dataPoint is a live metric. Seed-mode cycles still use full history unchanged.

Side-effect: if <5 live points exist after filtering, MIN_HISTORY_POINTS guard
naturally suppresses detection — correct behavior while live data accumulates.

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

Author: theo-learner

src/lib/detection-pipeline.ts

@@ -49,7 +49,14 @@ export async function runDetectionPipeline(
   balances?: { batcherBalanceEth?: number; proposerBalanceEth?: number; challengerBalanceEth?: number }
 ): Promise<DetectionResult> {
   const history = await getRecentMetrics();
-  const anomalies = detectAnomalies(dataPoint, history, balances);
+
+  // When the current dataPoint is a live metric (no seedTtlExpiry), exclude seed data
+  // from history to prevent stale spike scenario data from triggering false anomalies.
+  const filteredHistory = dataPoint.seedTtlExpiry
+    ? history
+    : history.filter(p => !p.seedTtlExpiry);
+
+  const anomalies = detectAnomalies(dataPoint, filteredHistory, balances);
 
   if (anomalies.length > 0) {
     logger.info(`[Detection] ${anomalies.length} anomalies detected`);