Weft is an early research thread around local inference scheduling on Apple Silicon. It is watching upstream correctness work before the first public artifact lands.
The shape is simple: keep tenants honest under load, make measurements easy to reproduce, and avoid pretending that local-first means single-user forever.