FINAL INDEPENDENT ADVERSARIAL MANUSCRIPT REVIEW
25 October 2026. Manuscript read in full through Section 9.
Reviewer: dream_adversary. Local CPU only; no external API writes.

VERDICT: PASS, no mathematical blocker identified.
This is an internal proof audit, not independent external validation or an
exhaustive literature-priority investigation. The main reference-process
contribution is substantive enough to justify the focused manuscript, provided
the present distinction from individual integer trajectories remains explicit.

1. CORE LAW AND COVARIANCES
I independently reconstructed the four-case defective base generating function,
its convolution with the outgoing geometric reserve, all four displayed raw
moments, and the genuine-change marginal moments. These agree with Sections
3--4. The conditional endpoint independence is correctly limited to revealed
types, not future types. Positive charges at actual canonical type changes
follow from strict ultrametric separation, not from a generic pin-distance
inequality applied without conditioning.

For the covariance proof, I independently obtained
 c_(r,s)=2-4*2^(-r)-6*2^(-r-s),
 C_h=sum_a pi_a^h c_a,
 K_h=2/(2^(h+1)-1)^2,
 gamma_h=C_h-C_(h+1)+2K_h-K_(h-1)-K_(h+1).
They simplify to the displayed negative kernel. In the K_h proof the
same-type event contributes conditional variance 2 even though its conditional
R_0 mean is shifted by h*B_a; that shift multiplies a zero conditional mean
of R_h-2. In the C_h proof, a first change before h determines R_0 from
types of indices less than h, so its contribution against B_h-4 vanishes.
The telescoped covariance sum is -16/9 and long-run variance is 4.

2. PROCESS LIMIT AND GENUINE-CHANGE CLOCK
The cumulative identity Y_m=Z_m+R_m-R_0, the endpoint union bound, and the
functional CLT argument are valid. No independence of endpoints and Z_m is
needed. The joint finite-dependence covariance matrix is
 [[4,8/27],[8/27,152/1323]].
Applying the joint functional limit at tau_n/n -> 9/8 and the exact equality
N_(tau_n)=n gives variance 807/196. This is not the conditional marginal
variance 25/4, and the manuscript correctly distinguishes them. Its elementary
blocking justification for the finite-dependence invariance principle is sound.

Minor clarity edits communicated to the author:
 - The negative-binomial formula in (20) is for m>=1; m=0 is separately the
   deterministic zero sum.
 - The vector FCLT initially sums B_0,...,B_(m-1), while Z_m sums B_1,...,B_m.
   Their difference B_m-B_0 is uniformly negligible up to Cn after division by
   sqrt(n), by exponential tails. State this explicitly at the clock step.
These are easy presentation clarifications, not failures of the conclusions.

3. FINITE-SOURCE TRANSFER
Let Q=sum_(i=0)^m B_i. When Q+R_m<H, an odd source class mod 2^H fixes all
revealed blocks and the exact terminal valuation. Backwards reserve recovery
then fixes every charge. The condition is safely strict; a valuation equal to
the number of remaining bits must remain unresolved. Terminal R_m is geometric
independently of the revealed types, so Q+R_m is exactly a sum of 2m+3
positive geometric variables. A maximal coupling of source residues and Haar
residues gives epsilon+delta, not 2*epsilon+delta: the failure event is measured
under the Haar component and every resolved matching cylinder fixes the tuple.

The sharpened discrepancy suggested by the certifier is correct:
 K=2^(H-1), s=N mod K,
 delta=s*(K-s)/(K*N) <= K/(4*N).
Indeed s classes have (floor(N/K)+1)/N mass and K-s have floor(N/K)/N mass;
their half-L1 discrepancy from 1/K is the stated expression. The original
weaker bound is also valid.

For theta in (0,1), E[2^(theta*G)]=2^theta/(2-2^theta); Markov therefore gives
 epsilon <= 2^(-theta*H)*(2^theta/(2-2^theta))^(2m+3).
No unproved independence beyond the established type/reserve coding is used.

4. GROWING-WINDOW COROLLARY (PROPOSED AUTHOR ADDITION)
The claimed growing-window limit is valid. If m_N -> infinity and
limsup m_N/log2(N)<1/4, choose a fixed c between that limsup and 1/4 and
choose alpha in (4c,1). Set H=floor(alpha*log2(N)). For sufficiently large N,
m_N<=c*log2(N). The base-2 logarithm of the Chernoff bound, divided by log2(N),
is at most
 -theta*alpha + 2c*log2(2^theta/(2-2^theta)) + o(1).
Its derivative at theta=0 is -alpha+4c<0, so some fixed positive theta makes
it negative. Consequently epsilon ->0 and delta=O(N^(alpha-1))->0.

TV contraction under the (N-dependent) measurable centered-sum map now gives
 (Y_(m_N)-4m_N)/sqrt(m_N) => Normal(0,4)
for a source uniform on the first N positive odds. Undefined tuples carry a
cemetery symbol whose mass tends to zero by the same TV bound. This is a
controlled joint source/time limit; it gives neither a fixed-integer theorem
nor transport under first-descent conditioning. The manuscript's current
opening phrase 'source-size limit taken before any infinite-time limit' should
be updated if this joint-limit corollary is inserted. The appropriate assertion
is that time is fixed or grows only within the proved precision budget.

5. SCOPE AND PRIORITY
Classical exact cylinders, IID geometric exponents, generic coboundary CLTs,
and ordinary finite-source residue coupling are background tools. The novelty
focus should remain the adaptive selected reserve observable, its exact charge
law, explicit all-lag anticorrelation, and the selected-clock calibration.
Prior rStation dictionary and endpoint-distribution results are not superseded
by these Haar laws. The manuscript already avoids such claims. The positive
integer transfer strengthens the concrete applicability of the observable but
does not improve a Collatz descent theorem. No universal growth-debt ceiling,
no cycle exclusion, and no almost-sure-to-every-integer inference was found in
the manuscript. Keep all those qualifications in the submitted version.

Review outcome: PASS; no unresolved blocker. Parent may incorporate the minor
clarifications and the independently verified growing-window corollary.
