Benchmarking ngram-simple, ngram-mod, and draft-target speculative decoding on synthetic multi-turn edit workloads. T1 creates the artifact; T2-T4 make small edits and re-emit the full result.