SparseLinearDRLearner doesn't shrink point estimates to zero at all

Hi, thanks for the amazing library!

I have some problems while experimenting with the SparseLinearDRLearner, and I hope I could borrow your mind on them.

1. In my case, I have about 10000 samples and about 50 features in the CATE function, and I wish to obtain a sparse version of the function by using SparseLinearDRLearner. However, it seems that the SparseLinearDRLearner won't shrink any point estimates to zero at all regardless of the value of alpha (0.0001 to 10000). Could you explain why this may happen?

2. When comparing the results of SparseLinearDRLearner and LinearDRLearner, I found that the point then p-values and confidence intervals presented in the summary tables can be very different for the same features. For example, for some feature A, the result of the SparseLinearDRLearner states that the p-value is 0.001 whereas that of the LinearDRLearner is 0.6, hence giving a totally different insight. How would you interpret this dichotomy? Thanks!

Start by reproducing the reported behavior with SparseLinearDRLearner and LinearDRLearner using roughly 10,000 samples, 50 features, and the stated alpha range. Compare their point estimates, p-values, and confidence intervals, then inspect the learner implementation and documentation to determine whether the observed differences are expected. Done means providing a clear explanation or identifying a reproducible defect.