I've typically used 95 because that seems like a suitably conservative number. But I also do some shorter runs to see where those end up, because the longer runs have fewer years available to run the simulations. If you use an extreme old age number, thinking you're being conservative, you may actually end up being less conservative due to the smaller sample size.

Assuming entitlements and my pension don't get changed too much by the pols (a big assumption), those sources will cover basic expenses by the time I reach my early sixties. So I'm also doing runs that end when I hit 62. This gives me a different perspective - kind of a worst-case scenario - will my savings last at least until I'm able to go on the govt. dole? This makes for a much larger sample size and allows me to see the impact of the dot.com crash in '00.

This. The longest stock market series data I know of is the Schiller dataset which goes back to 1871. So 145 years of data. The number of time segments you can test is (Number of Years of Data) - (Length of Retirement) + 1. That means someone FIREing at 55, planning to live to 95 has a 30 year retirement, and 115 30 year periods to test their withdrawal rate on but a 30 year old planning to live to 110 has an 80 year retirement and only 65 80 year periods to test their withdrawal rate on.

The other problem with extremely long retirement windows in the type of calculation cFIREsim does is that they're removing the retirement years closest to the present, which, if you think the stock market may behave differently today than it did back in the 19th century, is the very last thing you want to do. A 30 year retirement window can calculate the outcome of retirements from 1871-1985. For an 80 year retirement, the most recent window in your analysis is a retirement staring in 1945.

I'd rather look at more, shorter time windows, and maybe consider the cases where I wouldn't be out of money, but am well below my starting balance after 30 years as "likely failures" instead of throwing out so much of the data.