sim-lib-compute-auto

Automatic compute-site selection for SIM tensors.

AutoTensorExecutor accepts measured profile records captured through a bounded benchmark harness. Profiles record adapter, driver, backend, limits, transfer and allocation evidence, tile sizes, operation sample distributions, thermal and power context, and provenance. They persist through a caller-supplied Table or Dir value with bounded keys and record sizes; the auto library never opens a host path.

Automatic placement stays conservative: absent, stale, incompatible, or inconclusive measured evidence routes to the published CPU tensor executor from sim-numbers. Fresh compatible evidence can select the modeled resident provider, and each execute/flush records provider choice, materialization bytes, and synchronization counts in the routing ledger.