Frozen2026-11-02 before the finite capacity benchmark.

Avoid extrapolating a limiting support ceiling from short trajectories.
Instead compute actual finite selected ordered-type TV with rigorous
normalizer brackets, to make the abstract target inspectable.

N=2^L initial odd sources, L8,10,12,14; n1..8 blocks. All rows retained.
Enumerate literal canonical blocks, keeping every actual no-descent source,
including1. Model word mass is2^-Q only if every coefficient prefix stays
positive; otherwise0. Exact finite-cost DP at H96 yields p_lower.
The unconditional negative-binomial tail
 Pr(Q_n>=96)=2^-95*sum_{k=0}^{2n-1}binom(95,k)
gives p_upper=min(1,p_lower+tail). This is a rigorous bracket for p_n,
not a statistical confidence interval. For actual atom masses a_w,
TV=1-sum_w min(a_w,2^-Q(w)/p_n), over model-selected support words.
Monotonicity brackets TV by substituting p_lower,p_upper.

Verify the atom-overlap formula independently on a finite normalized
surrogate distribution using a full l1 calculation. Report selection-rate
ratios, support size and exact rational TV brackets. No inference of a
new asymptotic onset or confirmation of the limiting ceiling from these
small examples. Cap60seconds, one CPU process, no external compute.

Phase2 amendment, before coarse comparison: retain ALL32 rows and compare
the joint totals(R,S) with the full ordered type word. Use exact selected
type-count convolution at the same H96 and a conservative omitted-tail
overlap allowance. Verify the data-processing inequality with the resulting
brackets; no extrapolation that a coarse record remains adequate forever.
