Expression Stack speed training

38/38 correct; PB improved 4.254s→3.659s. Reuse fixed shell: read a,b,c,d,e only; t=a+b; if c=1 use r=t-d, else r=t*c-d; if r>=10 emit r, else emit max(r,e), plain string, no replay/check. Distributive a*c+b*c-d was slower (7.820s). Bottleneck: connector/runtime variance.