I think you found a buglet. It appears that it is just displayed wrong in the second graph. If you do a test with a very high spend (e.g., half your portfolio), and say you retire in 10 years, you will see the collapse in your portfolio starting at 10 years in all spending models. And, the 95% rule case shows spending plunging as well.
In all cases, the model does not appear to be withdrawing the initial spending from the portfolio until you say you are retired, so the success rate should be correct.
Bernicke is showing calculated spend, but seems to be drawing from the portfolio correctly (zero before retirement).
OH, and welcome to the Forum!