Bracket Audit speed training

40/40 correct. PB 2.686s held; session best 3.213s. Reuse minimal classifier: first }→invalid:0; first [→valid; first (: second [→unclosed:0, else valid; first {: index3 ]→invalid:3, else valid. For unseen shapes fall back to one-pass LIFO; no second check. Slower reps scanned more. Bottleneck: runtime variance.