Order of columns and model accuracy (AUC, F1) on XGBRegressor

**Context**
I have a very large number of input variables (100s), but only a handful (~50) matters.
I use XGBRegressor to build classification models (xgboost 0.90)
The model's accuracy is quite good (AUC 0.875, F1 0.846)

**Problem**

1. Rebuilding the model using only the important features gives a lower accuracy (AUC 0.872, F1 0.750) than the original model (getting the list of 39 features using `my_model_xgboost.get_booster().get_score`)
2. Rebuilding the model using Shap's most important features gives the same accuracy as the original model. Shap outputs the same 39 columns as above, but in a different order

This does not reproduce on tiny datasets, like the Iris one.

Given that the drop in accuracy is quite significant, I would consider this a bug in the optimisation step of XGBoost. I understand that with equal importance XGBoost would pick the first one, but this does not seem like an equal importance step.