Actually, if one looks at the rshares generated during voting, they are very large numbers, so I suspect the convergence happens very rapidly.
I also wonder if the code will use some approximating function for such large numbers, as happens with the square-root function in curation rewards calculations.