2.4. Hooks define when asynchronous work is finished🔗

A host call can return before a GPU, accelerator, or foreign task has completed. Stopping the clock at that return would measure submission latency rather than completed work. SpanHooks lets an adapter define the boundary without putting CUDA, PyTorch, or another runtime into the core package:

prepare
  start clock
    run action
    completeTiming
  stop clock
enrich metadata

prepare creates adapter state before timing. completeTiming may synchronize the selected runtime before the stop sample. enrich runs outside the measured interval and can attach allocator or device counters. The metadata should name the timing convention because a synchronized device span and an unsynchronized host-launch span answer different questions.

The hook does not expose the kernels or operators inside a foreign call. It gives the outer Lean application an honest completion boundary. PyTorch Profiler, Nsight, Perfetto, or another runtime-specific tool remains the right instrument when the question lies below that boundary.