A DESCRIPTION OF THE PROBLEM :
My apologies in advance for what is probably my mistake. I'm looking at JRE1.8.0-241 javax.vecmath.Matrix4d (revision 127, 2008-02-28) at line 915+.

914 ww = 0.25*(1.0 + tmp_rot[0] + tmp_rot[4] + tmp_rot[8]);
915 if(!((ww<0?-ww:ww) < 1.0e-30)) {
916 q1.w = (float)Math.sqrt(ww);
917 ww = 0.25/q1.w;

at line 915 ww can be negative, which means on 916 sqrt(negative) returns NaN. Oops?

STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
see test case below

EXPECTED VERSUS ACTUAL BEHAVIOR :
EXPECTED -
non-NaN values in B.
ACTUAL -
B.* is NaN

---------- BEGIN SOURCE ----------

public class Matrix4dTest {
@Test
public void testQuatFail() {
Matrix4d A = new Matrix4d(
6.123233995736766E-17, 6.123233995736766E-17, 1.0, 44.5,
-6.123233995736766E-17, -1.0, 6.123233995736766E-17, 1.3949339365687892E-16,
1.0, -6.123233995736766E-17, -6.123233995736766E-17, 61.967099999999995,
0.0, 0.0, 0.0, 1.0);
Quat4d B= new Quat4d();
A.get(B);
assert(!Double.isNaN(B.x));
assert(!Double.isNaN(B.y));
assert(!Double.isNaN(B.z));
assert(!Double.isNaN(B.w));
}
}
---------- END SOURCE ----------

CUSTOMER SUBMITTED WORKAROUND :
        ww = 0.25*(1.0 + tmp_rot[0] + tmp_rot[4] + tmp_rot[8]);
        double wAbs = Math.abs(ww);
        if(wAbs >= 1.0e-30) {
          q1.w = Math.sqrt(wAbs);

FREQUENCY : always