Request-level chunk policy. Execution consumes the selected size; model backends can account for their temporary workspace without changing admission.
Request-level chunk policy. Execution consumes the selected size; model backends can account for their temporary workspace without changing admission.