Strip heavy fields off a KRLS fit — slim

Removes large n x n (or n x m) intermediates from a fitted krls_rr object so it serialises (saveRDS) much smaller. Two modes:

Usage

slim_krls(fit, keep_predict = TRUE)

fit: A krls_rr fit.
keep_predict: Logical. If TRUE (default), keep enough state for predict() to run.

A krls_rr object with heavy fields removed and $slimmed = TRUE.

keep_predict = TRUE (default): keeps everything predict.krls_rr() needs (coeffs, sigma / sigma_vec, lambda, the standardisation stats, factor bookkeeping, and Nystrom landmarks where applicable). Drops the K, eigenvector matrices V / Vsq, projected response Vty, and vcov.c.
keep_predict = FALSE: keeps only inspection-grade summary fields (coeffs, R2, Looe, avgderivatives, var.avgderivatives, sigma, sigma_vec, lambda, Neffective). predict() will refuse to run.

The returned object carries the flag $slimmed = TRUE. unslim_krls() is a no-op (heavy intermediates can only be rebuilt by refitting).