We observe a different behavior between jdk8u and jdk9+. eg.:

% cat Test.java
import java.time.Duration;

public class Test {

     public static void main(String []args){
        System.out.println(Duration.ofNanos(-57282645800L).toMillis());
     }
}

% JavaSE8/jdk1.8/bin/java -Xmx128M -Xms16M Test
-57283
% JavaSE11/jdk-11/bin/java -Xmx128M -Xms16M Test
-57282

The answer depends on the intermediate representation of a Duration. for Duration.ofNanos(-57282645800L), its seconds = -58 and nanos = 717354200. JDK-8146747 introduces a change for java.time.Duration.toMillis(). The new toMillis() breaks this comment of nanos because tempNanos could be negative.

    /**
     * The number of nanoseconds in the duration, expressed as a fraction of the
     * number of seconds. This is always positive, and never exceeds 999,999,999.
     */
    private final int nanos;

I am not saying this is a bug, but the results are different from jdk8u and later JDKs.

If we only want to deal with Long.MIN_VALUE, why do we borrow a second for all negative durations?
 
       if (tempSeconds < 0) {
            // change the seconds and nano value to
            // handle Long.MIN_VALUE case
            tempSeconds = tempSeconds + 1;
            tempNanos = tempNanos - NANOS_PER_SECOND;
        }