Curie-temperature methods on RE-free measured anchors

Measured Curie temperatures for seven magnet anchors compared against two cheap prediction methods. rf_loo_pred_k is the BYU group's released random-forest model (github.com/msg-byu/ML-for-CurieTemp-Predictions, commit 80841ce4) with every row of the anchor's composition removed before training, so no value is scored from a model that saw the compound's own measured Tc. rows_in_hart_data / hart_data_tc_values report what the released DS1+DS2 data holds for that composition (their data encodes non-ferromagnetic elements as Tc = 0). tb2j_mc_tc_k is our own TB2J plus Monte Carlo prediction from the RE-Free Permanent Magnet Leaderboard work. Method controls (2026-09-24): their 67/33 recipe reproduces at MAE 74.5 K (R2 0.849); train-DS1 test-DS2 gives 90.5 K; leave-one-out on 100 random compositions of their own data gives MAE 61.7 K with bias -23.8 K.