In general, the observed keyword is used to define the likelihood terms (as opposed to the priors on the parameter terms), however, with DensityDist if using a dictionary, the observed keyword is used to pass kwargs to the likelihood function, any kwargs, observations, model parameters, constants…
Taking a look at the multiple possibilities to do exactly the same thing with DensityDist from Using a random variable as observed - #5 by OriolAbril might help understand how observed in DensityDist has nothing to do with observed keyword in all other density distributions