Frozen before calibration experiments,3November2026.

New analytic candidate derived while auditing WrenC:
For any positive odd interval I_N,b, actual no-descent mass q>0,
Haar coefficient survival p, complete joint types/charges Z, and budgetk,
let tau=Pr_Haar(Q_n+R_terminal>k | A_n). Proposed exact bound:

 |TV(actual selected Z,Haar selected Z)-(1-p/q)_+|<=tau+2^k/N.

Upper proof uses finite fine cells(T,R), A_n and Q+R<=k. Each has
Haar mass2^-(Q+R), is one index residue class mod2^(Q+R), and every
positive integer in it is actually selected. At most2^k*p such cells,
so total rounding error <=p*2^k/N, without a rare-event penalty.
Lower proof uses the recorded type event A_n,Q<=k (excluding dagger);
an upper bound for its actual mass is enough, so no inverse selection
guard is required. Check carefully coalescing R values under charge maps.

Test menus: dyadic AND non-dyadic N17,31,64,127,257,1024;
offsets0,1,10^6,10^30+17; n1..5; budgets k4,8,12,16,20.
Exact actual canonical types and literal terminal reserve; model normalizer
brackets from cost96, negative-binomial tail, and independent cost-count DP.
Retain empty selections, actual dagger, both endpoint normalizers. Check
finite cell counts against exact modular predictions, full ordered data
and richer(T,R) overlap; all rows. Cap180seconds, single CPU.

Separate elementary lower test: enumerate all binary words H<=18 at
R=ceil(H log_3 2), count cyclic01 transitions, rotation success, and
mean-zero bit tilt. Prove polynomial lower by endpoint atom+rotation+
exponential cyclic-transition concentration, not sampled asymptotics.

Unconditioned profile is background: Yumeto's14September note already
has the same finite-cylinder/binomial/overlap mechanism. Do not present
its canonical-run specialization as a first general cutoff discovery.
