root finding uses refinement?

[Ben Abbott]

I may have lost my mind, but I thought Octave used a script to  
implement roots() ... perhaps this has recently changed?

In any event, I'd like to know if any root refinement is used by the  
built-in function roots()?

I have an application in which small differences in roots can be  
important. I'd like to improve the accuracy of the root finder so  
that I may place a higher standard on pole multiplicity when finding  
the residues of a quotient of polynomials (using 'residue')

I checked a simple polynomial ...

r = roots([1 0 18 0 81]) / 1i
r =

   3.0000 + 0.0000i
  -3.0000 + 0.0000i
   3.0000 - 0.0000i
  -3.0000 - 0.0000i

octave:20> r - round(r)
ans =

  1.6920e-08 + 2.5814e-08i
  -1.6920e-08 + 2.5814e-08i
  -1.6920e-08 - 2.5814e-08i
  1.6920e-08 - 2.5814e-08i

octave:21> eps
ans =  2.2204e-16

The error in the solution for the roots is about 8 orders of  
magnitude larger than the numerical limit ... is this the simple  
indicative of the present state of root finding or might a more  
refinement method improve upon this?