Fix startup time calculation for restarted containers (#317)

For is_restart=true records, calculate time_to_ready_ms from running_at
instead of pending_at. pending_at is the pod's original creation time and
never resets across container restarts, causing crashlooping pods to report
wildly inflated startup times (e.g. 318 hours instead of 19 seconds).

This fixes the root cause identified in services PR #7552, allowing the
CASE/helper workarounds there to be removed.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: push-pull-deploy

internal/collector/pod_collector.go

@@ -602,13 +602,22 @@ func (c *PodCollector) trackStartupLifecycle(_, newPod *corev1.Pod) {
 		if newStatus.Ready && entry.runningAt != nil {
 			// Calculate durations
 			var timeToRunningMs, timeToReadyMs *int64
-			if entry.pendingAt != nil && entry.runningAt != nil {
-				ms := entry.runningAt.Sub(*entry.pendingAt).Milliseconds()
-				timeToRunningMs = &ms
-			}
-			if entry.pendingAt != nil {
-				ms := now.Sub(*entry.pendingAt).Milliseconds()
-				timeToReadyMs = &ms
+			if entry.isRestart {
+				// For restarts, pendingAt is the pod's original creation time (never resets),
+				// so measure from runningAt to get the actual per-restart duration.
+				if entry.runningAt != nil {
+					ms := now.Sub(*entry.runningAt).Milliseconds()
+					timeToReadyMs = &ms
+				}
+			} else {
+				if entry.pendingAt != nil && entry.runningAt != nil {
+					ms := entry.runningAt.Sub(*entry.pendingAt).Milliseconds()
+					timeToRunningMs = &ms
+				}
+				if entry.pendingAt != nil {
+					ms := now.Sub(*entry.pendingAt).Milliseconds()
+					timeToReadyMs = &ms
+				}
 			}
 
 			c.emitStartupLifecycleEvent(entry, &now, timeToRunningMs, timeToReadyMs)
@@ -729,14 +738,23 @@ func (c *PodCollector) snapshotStartupLifecycles(pod *corev1.Pod) {
 
 		var timeToRunningMs, timeToReadyMs *int64
 
-		ms := runningAt.Sub(pendingAt).Milliseconds()
-		if ms > 0 {
-			timeToRunningMs = &ms
-		}
+		if entry.isRestart {
+			// For restarts, pendingAt is the pod's original creation time (never resets),
+			// so measure from runningAt to get the actual per-restart duration.
+			ms := readyAt.Sub(runningAt).Milliseconds()
+			if ms > 0 {
+				timeToReadyMs = &ms
+			}
+		} else {
+			ms := runningAt.Sub(pendingAt).Milliseconds()
+			if ms > 0 {
+				timeToRunningMs = &ms
+			}
 
-		ms = readyAt.Sub(pendingAt).Milliseconds()
-		if ms > 0 {
-			timeToReadyMs = &ms
+			ms = readyAt.Sub(pendingAt).Milliseconds()
+			if ms > 0 {
+				timeToReadyMs = &ms
+			}
 		}
 
 		c.emitStartupLifecycleEvent(entry, readyAt, timeToRunningMs, timeToReadyMs)