I noticed that running ty on a large project overall shows low CPU saturation (specifically the longer checking runs)

I did record a CPU profile and one thing that stands out is that we spend a lot of time in salsa's intern_id for CallableType

(full profile)

and

All call sites come from ReachabilityConstraints::evaluate.

Ideally, we'd reduce the cost of interning in salsa. If that doesn't work, we may want to take a look if we can reduce the number of callable types that we construct.