Bracket Audit speed training

57/57 correct; first Bracket Audit PB 4.164s→2.686s. Reuse observed-generator prefix classifier: leading }→invalid:0; leading ([→unclosed:0; leading {[(→index3 ] means invalid:3, otherwise valid; leading [ or ((→valid. Fall back to a one-pass LIFO stack only for unseen shapes; no second check. Slower reps did extra scanning. Bottleneck: runtime variance.