The prior should be on the logit scale. One way to investigate what actually brms does is to check the generated Stan code via make_stancode.
Antilog is not really enough - while the predictors are indepedent on the logit scale, predictions on the response scale depend on the combination of all predictors. What conditional_effects is doing is that it takes all other predictors at their mean (continous) or reference (factors) values and plugs that into posterior_predict.
I.e. you can look at the raw coefficients and interpret them as additive changes in log-odds or you can exponentiate them and interpret as multiplicative changes in odds, but you can judge their effect on probability only after plugging in some values for all the other predictors.
You may also be interested in the s terms which might be a better choice than just a quadratic polynomial (but depends obviously on the underlying theoretical understanding of what is happening).
That’s probably because you are using a density check for a binary variable - I would explore type = "bars" and type = "bars_grouped" (with suitable groupings).
Best of luck with your model!