Tabular Data - Continuous Variable Model Training and metric selection

I have three basic questions.
Context: I am predicting continuous variable like Rossmann example except that my variable is auto regressed (dependent on previous day sale).

Q1: I do not want to take log of my target variable ; because if i do so I can not optimize for R^square or MSE. Reason being when I being my predictions back to exp(Y-prediction) I get huge error. So in that case should I normalize my other continuous variables specially previous day sale etc.

Q2. If I do not use take log of target variable then my model performance is really bad; I want to know if there is a way to over sample my data since I have very limited data points?

Q3: Lastly Can I cap my target variable at group level e.g.sales in Tx <x MM and sales in RI<y MM etc.

appreciate your comments and response.