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

bit-serial learned reducer (L=256): tiers 1-7 exact on local eval

initial commit