Part of epic #278. Port the #51 best-iterate safeguard (keep_best, snapshot_params/restore, final_ll semantics, the share_comps/do_reject disables) to AMICAMLXNG. Newton fallback overshoots were keep_best's original motivation; the #264 gate measured zero overshoot on sample data, so this is robustness for harder data, not a known live defect.
Part of epic #278. Port the #51 best-iterate safeguard (keep_best, snapshot_params/restore, final_ll semantics, the share_comps/do_reject disables) to AMICAMLXNG. Newton fallback overshoots were keep_best's original motivation; the #264 gate measured zero overshoot on sample data, so this is robustness for harder data, not a known live defect.