Runs a small, fast subset of the corpus single-ref and fails if any instance is infeasible or exceeds a generous gap-to-BKS ceiling. Intended to run on every CI push as a cheap quality guard; the labeled-PR A/B job is the precise gate.
mix bench.smoke
mix bench.smoke --budget 10 --ceiling 0.05Options: --budget Per-instance budget in seconds (default: 10) --ceiling Max allowed gap-to-BKS as a fraction (default: 0.08 = 8%)