According to Chenoweth, the number refers to peak, not cumulative participation. She also says 3.5% is not absolute – even non-violent campaigns can succeed with less participation, according to her 2020 update to the rule.

That's the opposite of what her update said (well, it's rather misleading). Her update noted cases where nonviolence failed even when they beat 3.5% - including one case that achieved 6% participation. She did note that most successful attempts didn't need to reach 3.5%, but also that reaching that is no longer a guarantee.

Her original research only went to 2006, there's been a few recent cases which broke the rule. Like she said in her update, history isn't necessarily a predictor of future results. I think there are also some very recent cases like Nepal where 95% of the movement is nonviolent, but violence at the very end of the movement tips the scale. (IIRC something similar happened with the Iranian revolution, though the results of that were decidedly undemocratic in the long run). There's some nuance with Nepal as well- the organizers did not choose to go for violence, it was largely an unplanned mob reaction.

Based on the totality of her research (which is publicly accessible and based on publicly accessible data), I still think nonviolence is more likely to achieve success than violence, but it really annoys me when articles like this one overstate the effects. It makes it really easy to tear apart the argument.