THEOREM LEAD — EXACT SELECTED-REFILL LAW AND SERIAL DEPENDENCE
25 October2026. CPU-only bounded checks. No API writes/GPU/paid/SYMBIOSE.

SCOPE AND DEFINITIONS
All probability statements below concern normalized Haar measure on the odd
2-adic integers and the canonical block clock. They are NOT claims for every
positive integer. Positive integers form a countable Haar-null set.
Remove the countable exceptional pins and their finite preimages, including
states at which a branch/block has an infinite valuation. This leaves an
invariant full-measure set on which every block and every reserve used below
is finite. None of these null-set removals changes a finite probability law.
Let T_i=(r_i,s_i), B_i=r_i+s_i and alpha_i=alpha_(r_i,s_i). The canonical
block map takes odd state X_i to X_(i+1). Let
 R_i=V_(alpha_i)(X_(i+1))
be the outgoing reserve. This is finite almost surely. At the next interface,
 J_i=V_(alpha_(i+1))(X_(i+1))-V_(alpha_i)(X_(i+1))
is zero for identical types and strictly positive for a changed canonicaltype.
The old deterministic reserve identities and pin-separation formula are input
lemmas, not new probability assumptions.

1. BERNOULLI FULL-BRANCH CODING
A prescribed canonicalblock(r,s) is the exact odd exponent word(1^(r-1),s+1),
with S=r+s. Its source cylinder has Haar mass2^-S and the affine block maps
that cylinder bijectively onto the full odd2adic space, preserving normalized
conditional Haar measure. Therefore finite successive type words have mass
product_i2^-(r_i+s_i): the T_i are IID with pi_(r,s)=2^-(r+s).
Also X_(i+1) is Haar and independent of T_0,...,T_i. For each fixed odd pin,
the valuation of its distance from a Haar odd state has distribution
 P(R_i=g | T_0,...,T_i)=2^-g, g>=1.
Thus ER_i=2,VarR_i=2, independently of its OWN type. It is not independent
of future types: its large values encode the upcoming repeated word.
This is standard exact2adic coding, not an independence claim about a
deterministically chosen positive starting integer.

2. COMPLETE SEPARATOR POLYNOMIAL
For fixed type i=(r,s), let t(i,j)=v2(alpha_i-alpha_j), j!=i. The explicit
canonical formula is min(r,u) for u!=r, and r+min(s,v) for u=r,v!=s.
Summing the IID weights gives the exact identity
 sum_(j!=i) pi_j z^t(i,j) =sum_(h=1..B_i)2^-h z^h.                 (A)
One direct proof: for1<=h<r, the termsu=h contribute2^-h. At h=r, allu>r
contribute sum_(u>r)2^-u=2^-r. Forh=r+a,1<=a<s, termsu=r,v=a contribute
2^-(r+a). At h=r+s, termsu=r,v>s contribute2^-r*sum_(v>s)2^-v=2^-(r+s).
The omitted identicaltype has masspi_i=2^-B_i. This also describes the
truncated-geometric separator conditional on a genuine change.

At a change i->j, the next incoming reserve is B_j+R_(i+1), so
 J_i=B_j-t(i,j)+G,
where conditional on(i,j), G=R_(i+1) has law2^-g on g>=1. This G is NOT
independent of all future types or subsequent refill variables.

3. EXACT MARGINAL REFILL LAW
P(J=0)=sum_i pi_i²=1/9. Forh>=1,
 P(J=h)=h(3h-1)/(18*2^h),
 P(J>=h)=(3h²+5h+8)/(9*2^h).                                   (B)
Derivation: in the mixture at a change, let l=B_j-t(i,j)>=0. Using(A),
the unnormalized mass of shiftl is
 sum_(j: B_j>=l+1)pi_j*2^-(B_j-l)=(3l+1)/(9*2^l).
Here the l0 sum starts at B_j2; all B_j=s have multiplicitys-1. Convolving
with P(G=g)=2^-g gives the stated quadratic-times-geometric coefficient.
The PGF is
 E z^J=1/9+4z(1+z)/[9(2-z)^3].
Consequently EJ4, EJ²212/9, VarJ68/9, EJ³172, EJ^4=13508/9.
At an actual change, the conditional law is h(3h-1)/(16*2^h), h>=1,
with mean9/2 and variance25/4. Those are the selected-event marginals,
not an IID model for the sequence of changes.

4. EXACT RESERVOIR COVARIANCE
Define K_h=Cov(R_0,R_h). For everyh>=0,
 K_h=2/(2^(h+1)-1)^2.                                           (C)
Forh>=1, if the initial type changes before indexh, R_0 is determined by
T_0,...,T_(h-1), while the conditional mean of R_h given those variables is2.
That case contributes zero covariance. On T_0=...=T_(h-1)=i,
R_0=(h-1)B_i+R_(h-1). If T_h!=i, R_(h-1)=t(i,T_h) is determined byT_h and
E(R_h|T_h)=2, again contributing zero. If T_h=i, then R_(h-1)=B_i+R_h,
and the surviving conditional covariance is Var(R_h|T_h=i)=2. Its probability
weight ispi_i^(h+1). Summing gives2sum_i pi_i^(h+1)=2/(2^(h+1)-1)^2.
At h0 this is simply VarR0=2.

5. EXACT FUTURE-COST COVARIANCE
Let C_h=Cov(R_0,B_h), h>=1. The same first-change decomposition leaves
 C_h=sum_i pi_i^h c_i,
where c_i=Cov(R_0,B_1 | T_0=i). For i=(r,s),
 c_(r,s)=2-2^(2-r)-3*2^(1-r-s).                                 (D)
Here is a complete computation. With independent Geometric(1/2) U,V,
 E[R_0 |T_0=(r,s),T_1=(u,v)]
 =min(r,u)+1_(u=r)min(s,v)+2*1_((u,v)=(r,s)).
Define b_r=Emin(r,U)=2-2^(1-r) and
 a_r=E[(U-2)min(r,U)]=2-(r+1)2^(1-r).
The latter follows from
Cov(U,1_(U>=h))=2^(1-h)(h-1) and summingh1..r.
Therefore
 c_(r,s)=a_r+2^-r[(r-2)b_s+a_s]+2*pi_(r,s)*(r+s-4),
which simplifies to(D). Summing the geometric series, withx=2^h,
 C_h=2/(x-1)^2-4/[(x-1)(2x-1)]-6/(2x-1)^2
    =-2(x-2)(3x-2)/[(x-1)^2(2x-1)^2].                           (E)
In particular C_1=0 and C_h<0 for h>=2.

6. REFILL COVARIANCES AND A DIRECT ADJACENT CROSS-CHECK
The exact coboundary identity is
 J_i=B_(i+1)+R_(i+1)-R_i.                                      (F)
Forh>=1, independence of B_1 from later reservoirs and C_1=0 imply
 gamma_h=Cov(J_0,J_h)
 =C_h-C_(h+1)+2K_h-K_(h-1)-K_(h+1).                            (G)
At h0, VarJ=VarB+2VarR-2K1=4+4-4/9=68/9.
The parent's independent symbolic simplification of(G), x=2^h, is
 gamma_h=-4x(36x²-34x+7)/[(x-1)(2x-1)^2(4x-1)^2]<0.
The numerator polynomial shifted atx1 is36y²+38y+9, so all lags are negative.
In particular gamma_1=-664/441. Absolute covariances decay geometrically.
Summing(G) telescopes:
 sum_(h>=1)gamma_h=C1+K1-K0=-16/9.
Thus VarJ+2sumgamma=4. An IID model with the correct marginal overstates
the asymptotic variance by(68/9)/4=17/9.

An independent adjacent-moment derivation avoids(C),(E),(G). Condition on
T0=i,T1=j,T2=k. Ifi=j orj=k, one refill is0. OtherwiseR0=tij,R1=tjk,
and ER2=2, giving
 E[J0J1 |i,j,k]=(B_j+tjk-tij)(B_k+2-tjk).
Summing the tripleweights yields6392/441, hence covariance-664/441.
One way to evaluate is sum overifirst, with A_j=sum_(i!=j)pi_i tij
=2-(B_j+2)pi_j. The remaining moments give
 E[J0J1]=14+4sum pi²-2sum pi³+E[(1-pi_T)c_T]
 =14+4/9-2/49+40/441=6392/441.
For detail, write s=B_j,p=pi_j. The summed i-factor is
(1-p)(s+tjk)-A_j=(s-2+2p)+(1-p)tjk.
Over k!=j, sum pi_k(B_k+2-tjk)=4. Moreover
sum pi_k*tjk*(B_k+2-tjk)=6+c_j+2p.
For this use sum pi_k*tjk*B_k=8+c_j-p*s*(s+2),
sum pi_k*tjk²=6-(s²+4s+6)p, and A_j above. Therefore the remaining
fixed-j expression is4s-2+4p-2p²+(1-p)c_j; averaging gives the result.
The supplied checker separately sums finite triple-type boxes in O(M^4),
and uses Cauchy-Schwarz to bracket the omitted moment by
sqrt((13508/9)*(1-P_box^3)). At M20 its rational bracket already certifies
strictly negative adjacent covariance.

7. SUMS, LLN, FCLT AND EXTREMES
Summing(F) gives
 sum_(i=0..m-1)J_i=sum_(i=1..m)B_i+R_m-R_0.                     (H)
The B_i are IID with mean4,variance4 and exponential tails. Reservoirs have
uniform geometric tails. Hence R_m/m->0 almost surely by Borel-Cantelli,
and max_(j<=m)R_j/sqrt(m)->0 in probability by a union bound.
It follows from IID SLLN and Donsker's theorem that refill sums have mean-rate4,
almost-sure LLN and FCLT with variance-rate4, despite the marginal variance68/9.
This is a coboundary argument; no assumption of independent J_i is used.

Also max_(i<m)J_i/log2m->1 almost surely. Upper bound: use(B), the union bound
at dyadicm, and Borel-Cantelli with threshold(1+epsilon)log2m. Lower bound:
disjoint IID pairs with T_(2l)=(1,1), next r>=2 and B_(2l+1)>=h forceJ_(2l)>=h.
Their probability is bounded below by a fixed positive multiple of h*2^-h
forh>=3. At h=(1-epsilon)log2m the probability of no such pair is at most
exp[-c*m^epsilon*logm], summable at dyadicm. Interpolate monotonic maxima.
This last extreme-value corollary is optional; the parent may omit if novelty
comparison suggests it adds distraction. No refinedloglog centering is claimed.

8. CHANGE-CLOCK AUDIT (PARENT / ADVERSARY RESULT)
Let I_i=1_(T_i!=T_(i+1)), p=8/9. The vector(B_i,I_i) is1-dependent with
long-run VarB4,VarI152/1323,Cov(B,I)=8/27:
VarI=8/81,Cov(I0,I1)=1/49-1/81; Cov(B0,I0)=Cov(B1,I0)=4/27.
At the nth actual change, random-time inversion of the joint FCLT and(H)
gives mean perchange9/2 and variance-rate
 [4+(9/2)^2*(152/1323)-9*(8/27)]/(8/9)=807/196.
The R boundary remains negligible at the random index via the maximum-reserve
bound and N_n/n->1/p. Stationary Palm indexing or beginning with the first
change after block0 has the same asymptotic result; only a finite initial
waiting segment differs. This is not the conditional marginal variance25/4.
I independently checked the constants and random-index/boundary argument.

FILES / VALIDATION
refill_law_check.py + REFILL_LAW_AUDIT.json:392 finite separator-law checks,
32 exact marginal/tail probability brackets at type-box cutoffs4,8,12,20,
exact PGF/moments, five independent adjacent-moment brackets, and covariance
constants. All finite arithmetic uses integers/Fraction; no random sampling.
The written proofs carry the universal probability statements.

NOVELTY / LIMITS
Bernoulli valuation coding, geometric moments and coboundary limit arguments
are classical. The selected-interface refill law, closed serial covariance,
and change-clock consequences are the proposed substantive synthesis, subject
to the parent's prior-art audit. No claim that these formulas solve Collatz,
bound every integer trajectory, establish RH, or are unprecedented without
comparison is authorized. Parent alone decides publication. All jobs finished.
