add honest limitation note (not exact on long chains; Neural GPU drift)

bit-serial learned reducer