Same Evidence, Different Target: Decoding How Diagnostic Evidence Bears on Causal Questions from Language-Model States