Can the user user features for train be used for test if the data is chronological

If we train a recommender system using Light FM and want to pass user features, how to do that in data that is time-dependent.

For evaluation, we divide the data into training and test set depending on date.

Item feature would be static so should be good.
But what about user features, such as average order value etc?
That would be different in the train and test dataset. And I would assume that won't work seeing the way we are creating the user feature.

Any advice on how to do that?

Also, how user features can be used? As in do we use it to get similar user.
Asking this because we have order history for the users and discounts that they have used. but we do not have same for train set(as the orders haven't been made).
So if we use it for similarity that would be fine.

So if you can give a little background on how these ffeatures are being used in the background, that would be very helpful.

No file, test, or entry point is named. Start by locating the LightFM documentation and examples for user and item features, then clarify how chronological train/test data should be handled and what feature use is supported. Done means the issue has concrete, documented guidance rather than an unresolved usage question.