Is this branching behavior expected?

[rezabayani]

In the branching example, after a MIP model is defined, an environment is created, reset, and env.step(action) is used for branching until done. I set expert_probability to 1.00 so that all actions are stored as training data. Here is code output for a model:
"Episode 1, 15 samples collected so far"
When I examine the observation. I notice two things.

1. The number of constraints in the first sample is different than the number of constraints in the rest of the samples. In my case, len(sample_observation.row_features)': array([4541., 4516., 4516., 4516., 4516., 4516., 4516., 4516., 4516.,4516., 4516., 4516., 4516., 4516., 4516.,]). Why is this happening? Shouldn't the number of constraints be increasing after each split?
2. Some actions are repeated. For example here, 'best_action': array([391., 513., 277., 61., 11., 149., 258., 323., 513., 277., 61., 149., 487., 323., 54.,] we see that actions 61, 277, 323 and 149 are taken twice. Is this expected behavior?
3. I noticed that in the presented model, the training data does not differentiate between where in the branching process the scores were stored. Is this right? What is the impact of expert_probability parameter?

I appreciate any feedback.