evals: tunable reasoning budget, so reasoning models score fairly

evals: benchmark candidate models for the t=0 web-search gate

eval harness + env-swappable frontier model