Code Review

This pull request introduces scikit-learn estimators (TabPFNClassifier and TabPFNRegressor) backed by a self-hosted TabPFN inference container over HTTP. The feedback suggests several improvements: implementing deepcopy to prevent crashes when using scikit-learn utilities like GridSearchCV due to the uncopyable httpx.Client, optimizing _to_jsonable to avoid redundant conversions when the input is already a list, adding defensive checks in fit to validate non-empty inputs and detect NaN values, and improving HTTP error handling to extract and raise detailed server error messages.

I think here we'd have the opportunity to cleanly implement sklearn interfaces, as in server, which makes sense now given kv cache semantics--- fit returns model_id, predict predicts.
Is there a reason to stick with single endpoint here? This was more a workaround for the market places?

otherwise LGTM

Two points:

1. model_id in the predict() signature is problematic, it invalidates the reason we (painfully) try to adhere to the sklearn estimator's interface.
2. I'm not sure I buy this should be a separate interface entirely from standard estimators. I think we should rather have a common base class.

Deferring to you on both points, but I think 1) is important.

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b261ff5. Configure here.}