J-space comparisons across open models

Six experiments extending the Verbalizable-Workspace paper on open models: temporal reach, training dynamics, cross-model transplant, capacity scaling, corpus dependence, MoE.