@Frank thanks you are right. when I rewrite body of udf whole example looks like:

import org.apache.spark.sql.functions.{udf, struct}

val reduceItems = (items: Row) => {
  // getting array of struct from struct
  val a = items.getAs[Seq[Row]]("second")
  // summing struct item in array
  a.map(_.getAs[Int]("navs")).reduce(_ + _)
}
val reduceItemsUdf = udf(reduceItems)

// passing struct of array of struct
h.select(reduceItemsUdf(struct("*")).as("r")).show()

and it works in spark 4, but I still do not know where was problem, why parameter can't be Seq?