When use ts_forecast with lgb, 2 more lags, results looks like just shift of series itself.

1. How exactly ts_forecast + lgb/ xgboost works? I looked it up in code seems it leverage https://hcrystalball.readthedocs.io/en/latest/, is there a documentation that introduce exactly what each parameters does and how those thing work together?
2. I am also trying to understand what 1) Optimize_for_horizon 2) lags 3) retrain_full means. For example, retrain_full I would expect the model use full train+test dataset, with tuned parameters and train the model at the end, is that the case? (Sorry if it's already in documentation, I did not find it)
3. I am seeing some strange behavior when lags>1 with loss MAPE, specifically the predicted series would looks just like a shift of original series, chart show below.