2026-11-02, frozen before running new count DP.

Prior accepted stopped_count_dp.py already counts old unknown cylinders
by 2D prefix-sum convolution. Do not rediscover that algorithm.
New task: adapt only the unfinished-run weight to the AppendixA rule.

For a selected j-block word with R ones,S zeros,Q=R+S<H, t=H-Q,
its old unfinished next-run prefixes are1^r0^(t-r), r1..t, one odd
H-bit cylinder per prefix. Let K_H=min{k:3^k>2^H}. The new decoder
retains such a prefix only when R+r>=K_H. Thus the new weight is
 w=max(0,min(t,H-S-K_H+1)).
Initial unfinished mass count H changes to H-K_H+1.
Terminal unknown count (one per completed selected type word) stays.

Controls: independently enumerate new decoder H8,10,12,14 andn1,2,3,5;
old DP vs original decoder, new DP vs modified decoder. Bound new<=old,
same selected count, all counts sum2^(H-1). No asymptotic claim.
Larger menu H120,240,480,800; n=floor(c(H-1)) for c1/5,2/9,7/30;
single sequential CPU run, total cap180seconds, no trajectory enumeration
at2^H. Use finite exact selected-cost mass only as a LOWER normalizer.
Record vacuous budgets; no cherry-picking a new global horizon.
