Windows PowerShell, 61
65 70 72 73 82
Evil alternative approach. It becomes apparent if you desperately search for patterns in the input.
Be sure to also check out Ventero's Ruby and Golfscript solutions which use the same technique which was sort of a joined work :-)
A little explanation might be in order for this, I guess. Suppose you have a number consisting of three digits. For convenience they are labeled a, b and c here with c being the largest. You then are going to subtract abc from cba.
If we picture that, we are going to get (c − a − 1), 9 and (10 + a - c) as the individual digits. This will always hold, no matter the input digits. This also means that the middle digit get no say in how long the sequence will be, simply because it always is
So basically the first and last digit of the sorted digits suffice in determining the length of the sequence. One can easily make a list with every conceivable combination of first and last digit (there are only 45) to search for patterns there.
As it turns out, there are at least two. Both play with the first few results of the different digits:
[6 5 4 3 1 2 3 4 5]. First, taking the difference between the largest and the smallest digit and subtracting one gives the appropriate index in the aforementioned list of results. Second, interpreting the smallest and largest digit as a single number and doing modulo 11 yields the same, except that one has to prepend another
6 to the list.