pub fn resolved() -> &'static TargetLeversExpand description
The process-wide resolution.
OnceLock-cached for the reason every lever it replaces was: these are
read per projection per layer per step, std::env::var allocates and takes
the process-wide environment lock (measured on GB10: 0.57 us
single-threaded, 5.76 us at 16 threads), and the route must be CONSTANT
across CUDA-graph replays — a per-call read could change the captured
launch set between capture and replay.