Eval harnesses

32 archived, timestamped evaluation runs; controlled A/B of raw model versus agent wrapper on identical queries to isolate framework overhead; time-to-first-token probes across query complexity.