Linear predictor metrics differ between `last_fit()` and `manual predict()`

[abichat]

Hi,

I’ve been experimenting with implementing a new metric based on the linear predictor (enabled by #551, thanks again for this work).

While testing it, I noticed a discrepancy between:

• the value returned by collect_metrics() after last_fit(), and
• the value obtained by manually calling my metric function on stats::predict(type = "lp").

I’m fairly confident (99.9%) that this is due to the fact that last_fit() ultimately calls parsnip::predict() with the default increasing = TRUE argument, which flips the direction of the linear predictor.

This behavior is not visible for royston_survival() since it is direction-invariant. However, for metrics that depend on the sign of the linear predictor, it’s a bit annoying that the result depends on whether you compute it via last_fit() or by hand.

Would it be possible to expose an option in new_linear_pred_survival_metric() to control the increasing argument used internally by last_fit()? Alternatively, is there a recommended way to ensure consistency between the two approaches?

Thanks again for the ongoing work on this infrastructure.