To test mass matrix adaptation I would
- initialize the chains using draws from the posterior, so that the initial warmup behavior is not adding extra variability
- after the last mass matrix change, run the steo size adaptation for a long time to reduce the variability in that part