sim-lib-compute-model

Modeled tensor compute provider for SIM.

The crate exports site/compute/model, an EvalFabric site backed by ModeledTensorExecutor. Accepted tensor requests run through the canonical sim-lib-numbers-tensor CPU semantics and return resident tensor storage owned by the modeled site. Resident storage supports deterministic readback faults and counters so recipes and tests can prove placement behavior without hardware. The resident pool is segmented by tile and binding limits, bounded by queue and arena budgets, and caches one synchronized materialization result.

The resident ODE adapter composes the published sim-numbers tensor ODE pipeline with the modeled tensor site. It batches fixed RK stages through bounded auto-flushed modeled submissions, keeps adaptive candidates modeled-resident, and limits adaptive host observation to scalar error decisions.