Using the selected normalizer in the actual-source bridge.
26 October2026. Proposed continuation of the25 October selected-consumer proof.
This is a conservative sufficient-window improvement, not a Collatz descent bound.

1. The same event and precision budget.

For n=m+1, A_n={S_j>0 for1<=j<=n}, with
S_j=sum_(i=0)^(j-1)[log(3/2)r_i-log2*s_i],
p_n=P(A_n). The joint charge/actual-first-descent coupling already proved
in SELECTED_CONSUMER_PROOF.txt gives conditional total variation<=2eta/p_n,
where eta=epsilon+delta+b, and
epsilon=P(sum of2m+3 positive Geom(1/2)>=H),
delta=s(K-s)/(KN), K=2^(H-1), s=N mod K,
b=min(N,(3^H-1)/2)/N.
The actual source is uniform on the firstN positive odds, and the requested
data retain the strict precision rule Q+R_m<H and small-source exception.
No marginal calibration or unbounded moment transfer is substituted.

2. A better lower bound for the normalizer.

Reuse the explicitly tilted q at t*, with lambda=M(t*) and E_q Y=0.
The unchanged exact change of measure is p_n=lambda^n D_n,
D_n=E_q[exp(-t*S_n);A_n]. Put epsilon0=1/128.
Rotation after a minimum and exchangeability imply
P_q(A_n,0<S_n<=n/128)>=P_q(0<S_n<=n/128)/n.
The centered finite-variance IID CLT makes the numerator tend to1/2.
Consequently for all sufficiently large n,
D_n>=exp(-n/128)/(4n), since0<t*<1.
By decreasing a positive constant c to cover the finitely many smaller n,
p_n>=c lambda^n exp(-n/128)/n for all n>=1.
This constant exists but is NOT made explicit by this CLT argument.

The certified lambda^4>1/2 gives lambda^-n<2^(n/4).
Also log2>5/8 (already its first atanh term gives2/3), so
exp(n/128)<2^(n/80). Hence
1/p_n<=C n 2^(21n/80).

3. Doubled actual-integer logarithmic window.

Let m<=log2(N)/10 and H=floor((3/5)log2N).
At Chernoff parameter2^theta=4/3 the positive-Geom mgf is2, so
epsilon<=2^(2m+3)(4/3)^(-H).
The exact integer inequalities
3^80<2^127, and4^80>2^33*3^80
give log2(3)<127/80 and log2(4/3)>33/80.
Thus, writing L=log2N, the exponents of both epsilon and b are
at most-19L/400+O(1); delta is O(N^(-2/5)).
After dividing by p_(m+1), the extra exponent is at most21L/800+O(1),
and there is the linear factor m+1. Therefore
dTV<=C'(m+1)N^(-17/800)=O((log N)N^(-17/800)).
In particular it is O(N^(-1/50)). Constants are not claimed practical,
sharp, or explicitly determined by the nonquantitative CLT lower bound.

The window m<=log2N/10 doubles yesterday's log2N/20 sufficient window.
For any m_N->infinity in that window, the same selected V_m/m limit d*
transfers in probability to actual first-descent-selected integer ensembles.
It does not imply convergence of their uncapped means or long-horizon descent.

4. Finite, independently evaluable normalizers.

The CLT constant need not be used for a finiteN calculation. A two-dimensional
integer DP counts all n-type words with totals R=sum r_i,S=sum s_i,
R+S<=H0, surviving every boundary. Its recursion is
C_n(R,S)=1_(3^R>2^(R+S)) sum_(r>=1,s>=1) C_(n-1)(R-r,S-s).
The convolution is exactly a2D prefix sum, costing O(n H0^2) operations.
Summing C_n(R,S)2^(-R-S) yields p_n^-<=p_n. The omitted mass is at most
P(sum of2n positive Geom>H0)=2^(-H0)sum_(k=0)^(2n-1)binom(H0,k).
This gives exact rational lower/upper bounds for p_n and an explicit
finite conditional-TV budget2eta/p_n^- whenever it is less than1.
The counting is for a normalizer only, not an estimate of any arithmetic
orbit average. Its small cases are checked against literal type-word enumeration.

At fixed n or m=0 handle trivial indexing separately; no assertion about
V_0/0 is made. Whole actual sources are retained; no chosen descendant is
resampled as a new uniform input.

Antecedents: classical parity/valuation coding and Cramer/rotation arguments,
the existing Collatz35 precision budget, and yesterday's selected-consumer
proof. The doubling is their explicit recombination with the same observable
and event; not a general new first-passage theorem. Compare Collatz13/18
before any priority claim. All new universal arguments await outside audit.
