feat(harness): hold recovery while a Grafana alert is firing, add legacy:busy host resolution

キャンペーン第 2 回のグレーゾーン 3 件を受けた変更。SLO 3 系列に現れない症状
（レプリケーション一時停止、メモリ圧迫）でも、受信中のアラートが解決するまで
recovered を出さないよう、AlertRouter が firing 中のアラートを保持し judge が
alerts_ok を条件に加える。legacy:busy は直近 2 分で最も多くリクエストを受けた
MIG インスタンス（Mimir topk）を指し、run+シナリオ単位でキャッシュして revert が
同じインスタンスに当たるようにする（ILB のハッシュで api の接続が 1 台に偏る）。

harness/gcp.go

@@ -60,6 +60,9 @@ func (a *App) gcpListInstances(ctx context.Context, prefix string) ([]string, er
 
 // resolveInstance turns "legacy:N" (N-th MIG instance by name) or a pet VM short name (db-primary -> selfheal-db-primary) into an instance.
 func (a *App) resolveInstance(ctx context.Context, ref string) (gceInstance, error) {
+	if ref == "legacy:busy" {
+		return a.resolveBusyLegacy(ctx)
+	}
 	if strings.HasPrefix(ref, "legacy:") {
 		n, _ := strconv.Atoi(strings.TrimPrefix(ref, "legacy:"))
 		l, err := a.gcpInstances(ctx, "legacy-")
@@ -170,3 +173,36 @@ func (a *App) gcpOp(ctx context.Context, op, instance, mig string, size int64) (
 
 func (a *App) region(zone string) string { return zone[:strings.LastIndex(zone, "-")] }
 func (a *App) regionOfMIG() string       { return env("GCP_REGION", "asia-northeast1") }
+
+// resolveBusyLegacy picks the MIG instance that served the most legacy-api requests over the last 2 minutes. The ILB
+// hashes the api pods' few keep-alive connections onto one instance (campaign 2, 2026-10-08: legacy:0 served 0 rps while
+// the delay was on it), so a fault must go where the traffic is. The choice is cached per run+scenario so the revert
+// hits the same instance after the traffic has moved away from it.
+func (a *App) resolveBusyLegacy(ctx context.Context) (gceInstance, error) {
+	key := ""
+	if cur := a.runs.current(); cur != nil {
+		key = cur.RunID + "/" + cur.ScenarioID
+	}
+	if key != "" {
+		if v, ok := a.busyLegacy.Load(key); ok {
+			return v.(gceInstance), nil
+		}
+	}
+	name, err := a.mimirTopInstance(ctx, `topk(1, sum by (instance) (rate(legacy_http_requests_total{route!="healthz"}[2m])))`)
+	if err != nil {
+		return gceInstance{}, fmt.Errorf("legacy:busy: %w", err)
+	}
+	l, err := a.gcpInstances(ctx, "legacy-")
+	if err != nil {
+		return gceInstance{}, err
+	}
+	for _, i := range l {
+		if i.Name == name {
+			if key != "" {
+				a.busyLegacy.Store(key, i)
+			}
+			return i, nil
+		}
+	}
+	return gceInstance{}, fmt.Errorf("legacy:busy: instance %q from Mimir not in the MIG (%d instances)", name, len(l))
+}