Correction: Reference preservation loss can oppose false-cutoff repair

Symmetric anchor preservation can oppose target correction on baseline-error rows: when an unfinished example is incorrectly predicted as finished, the target loss lowers end-of-turn probability while preservation pulls toward the incorrect anchor. Per-row scaling changes both terms together and cannot alter that within-row ratio. Treat this as a causal hypothesis until tested by disabling preservation only for affected unfinished rows, retaining it elsewhere, and evaluating with an independently loaded replay and a false-finished safety gate.