Trace length and length generalization across compression

Real runs from the paper: a two-layer transformer (about 6.7×10⁴ parameters, no positional encoding) trained on synthetic traces. EM0 is in-range exact match; EM1 and EM2 are two splits longer than any training input.