## 1. Notation and stated exact facts

Retained coefficients are scaled by three. Correction coefficients beta_e
remain unscaled; write w_e=3 beta_e, so every coordinate vertex has
weighted correction degree two. Let K be the set of N-attachment vertices,
m=|K|, and use the disjoint support counts

    Pi = stars + glues + N + Z + D,
    C = L3 + double-loop,
    R = Pi + A + C + X,

where A counts alternating directions and X mixed directions. In a loop
block B with source U={u,v}, put

    zeta_B(u)=Z_B(u)+D_B,
    z_B=zeta_B(u)+zeta_B(v),
    Psi(u)=sum_B zeta_B(u),
    Gamma(u)=sum_(private lines at u) s_l.

The support-safe normalization from the lower27 proof removes D when
at least one endpoint zeta vanishes. Afterwards D sources lie within K,
and Z occurs only at active endpoints. The identities and charges used
below are stated from the exact cell-law model:

    Gamma(u)-Psi(u)=-1/2,
    Pi >= 15+2m,
    t_U^G-z_(G,U)=Omega_G,
    I_e = -w_e + sum_(source(B)=e) z_B,
    sum_(G contains e) Omega_G = -w_e,
    2A+L3 >= 6-m,
    divergence(delta)=H-nu-w.

The generalized slot identity is replayed in all 30 orientations with
corrections left arbitrary. The group certificates are homogeneous cell
identities, so the displayed Omega law holds for every correction graph.
In particular some Omega is nonzero. Also

    sum_B z_B = sum_u Psi(u) = sum_e I_e + 6.                (1)

These are signed equalities; no argument below replaces them by positivity.

For a private line l with sum s and N contribution nu_l, its star vector is
l-s*1-nu_l. Set F(l)=glues(l)+2*stars(l)+2*N(l), f(u)=sum_(lines at u)F(l),
and pi(u)=f(u)/2+Z(u)+D(u)/2. Then sum pi=Pi exactly.

* Inactive vertices have f>=5. Every active vertex has two distinct N-lines,
  so f>=8. If Gamma is nonzero, f>=9.
* An F=4 N-line has no glue. An F=5 line with one N coefficient n has glue
  sum -n or -n/2. The corresponding forms are one glue/one star and three
  glues/no stars. These forms are derived from the V equation and replayed.
* At an active vertex with f=9 and Gamma nonzero, there are exactly two
  N columns, on lines of costs five and four. Four N plus the necessary
  star cost at least ten. Three N require a nonzero class with three of
  its distinct lines, costing at least twelve. Zero-zeta classes contribute
  cancelling pairs. No additional pure line has cost one or two.
* Equality Pi=15+2m fixes every vertex charge at its minimum. If some zeta
  is nonzero, its endpoints have one D of coefficient 1/2, no Z and f=8.
  D sources form a matching on K. All Z vanish. In the correction-matching
  branch this equality forces A>=3.

The last point is the full equality structure already proved at lower27.
For a fixed source U, distinct D blocks have disjoint h classes. With d
nonzero D columns on U, at least 2d N columns occur at each endpoint.
For m=2 this gives the useful bound

    Pi >= 12+4d+(d+z) = 12+5d+z,                          (2)

where z counts Z. The terms are inactive charge ten, at least one star
at each active center, at least 4d N columns, and d+z helpers. 

## 2. Three flow lemmas and a per-group charge

The six coordinate projections, valid for arbitrary corrections, are

    2 sum_B delta_B(1_W-1_V)
      = rho := sum_(source(B)={u,v})(Z_B(v)-Z_B(u))(e_u-e_v).

D cancels. If Z=0, a nonzero flow needs at least three blocks, since the
45 projected columns are nonzero and pairwise nonparallel. Every Z-free
nonzero-flow block needs two loop directions, so C>=2b in this case.
The loop-difference rows also show that no L3 implies delta=0.

**Single Z costs five loops.** Suppose only Z in block B0 is nonzero.
The right side rho is supported on its two source vertices, while the
projected column c_B0 vanishes there and has support on the other four.
After removing B0's flow, the remainder has support two or six. A single
other projected column has support four, so at least two other flow blocks
are needed. They are Z-free and cost two loops each; B0 costs at least one
L3 by its S rows. Thus C>=5, with disjoint charges. D does not affect this.

**Shared-source Z blocks.** If a vertex u belongs to every Z-block source
and rho(u) is nonzero, another, Z-free flow block must be nonzero because
all those source columns vanish at u. Hence

    C >= number of Z-blocks + 2.                           (3)

**Source-pair balance.** Suppose all nonzero endpoint zetas have source
U={u,v}, and delta=0. Define

    NT_U = sum_(splits separating u and v)(s(u)+s(v)).

Then

    NT_U = w_U-1.                                         (4)

The exact V/h elimination gives
H_U-nu_U-NT_U-(center_u+center_v)/2 as a linear combination of off-source
zetas. With centers one and off-source zetas zero this becomes
H_U-nu_U=1+NT_U; the flow equation proves (4). This argument permits
arbitrary zero-zeta N classes at vertices outside U. They cancel through
their h-side sums, not by assuming their individual coefficients zero.
All 15 source-pair identities are replayed with their off-source residuals.

If an endpoint carries a nonzero source-U class and a nonzero-sum private
line in a separating split, its two class N-lines lie in together splits
and each costs at least four. The separating line costs at least five.
Thus its f is at least 13. The nonzero-class hypothesis is required here.

**Per-group charge.** For Omega_G nonzero, every edge U of G not contained
in K costs one alternating direction in its array or one L3 in its block.
If t_U^G is nonzero, use the array. Otherwise z_B=-Omega_G is nonzero.
There must be exactly one active endpoint; the inactive endpoint has
zeta=0, normalized D=0, and the S-row difference forces an L3. The arrays
and loop blocks are distinct across all (G,U), so the charges add:

    A+L3 >= sum_(Omega_G != 0) #{U in G : U not contained in K}.  (5)

## 3. Five corrections require at least 24 retained directions

The signed degree census permits a double-star tree or C_4 plus an isolated
pair. Both have an edge of scaled weight exactly two. Let q be the number
of nonzero Omega groups. Since one matching can supply only three nonzero
edge coordinates, q>=2. We do not assume that an arbitrary C_4-plus-pair
Omega solution uses exactly two groups; kernel additions are allowed.

Every perfect matching has at most floor(m/2) edges wholly inside K.
Equation (5) gives the following cover.

* m>=5: Pi>=25. At m=4, Pi>=23 and A+L3>=2. At m=3,
  Pi>=21 and A+L3>=4. These give R>=25.
* m=2: if Pi>=20, (5) gives A+L3>=4 and R>=24. At Pi=19,
  the equality structure gives Z=0 and at most one D, of coefficient 1/2.
  Without D all 3q supported arrays are nonzero. With D at most one array
  can cancel, so A>=3q-1>=5. Again R>=24.
* m=1: if Pi>=18, (5) gives A+L3>=6. At Pi=17 equality
  permits neither D nor Z, so every zeta is zero. Effective private stars
  have support P0<=Pi and center mass one. Since P0<=17, at most six
  private lines have nonzero sums (each costs at least five and 2P0<=34).
  Every vertex is therefore simple, so H0 entries are in {0,1/2,1}.
  They cannot equal the weight-two correction edge. Flow is nonzero;
  with Z=0 it costs C>=6, while A>=3q>=6. This is more than needed.
* m=0: normalize D=Z=0. If private support P>=18, A>=3q>=6
  gives R>=24. If P<=17, the same simple-vertex argument gives nonzero
  flow and C>=6, so R>=15+6+6=27.

Thus e=5 implies R>=24 and W>=29. The proof uses only the simple-vertex
bound at P<=17, avoiding an extra dependency on the private support gap.


## 4. Matching: eliminate the general and one-vertex cases

It remains to exclude R=24 for matching corrections M. Lower27 already
excludes smaller retained support, so this is sufficient.

**N=0, directly.** Normalize D=Z=0. If flow vanishes, H=2M implies
P>=24 and A>=3. If flow is nonzero, C>=6, A>=3 and P>=15;
R=24 forces P=15,A=3,C=6,X=0. At P=15 every active private line
is tight, with sum -1/2 and positive stars. In the homogeneous matching
group equations, A=3 forces one column per M-array, each of coefficient
-2. The 64 possible array choices have eight D-compatible triples; each
occupies three blocks with distinct source pairs. A tight single-glue
private line produces two positive stars in distinct blocks with the same
source, so cannot occur. All private lines must instead be three-glue
lines, forcing one private split and six stars with six distinct sources.
They cannot fit the three alternating blocks. This directly gives R>=25;
the stronger N=0 width30 and two-partition theorems are unnecessary.

**All zetas zero with N>0.** Effective private support is at most Pi,
center masses stay one, H equals its effective value, and nu vanishes.
Also A>=3, so R=24 gives Pi<=21. The private H=2M lemma would
require effective support at least24. Hence nonzero flow forces C>=6,
and Pi>=17 gives R>=26. This case is excluded.

**m>=4.** At m>=5, Pi>=25. At m=4, Pi>=23; equality gives
A>=3, while Pi>=24 still needs A+L3>=1. Thus R>=25.

**m=1.** D=0, all Z attach at one vertex u, and some Z is present.
Let z be their count. The two matching edges avoiding u stay nonzero,
so A>=2. For z=1, Pi>=18 and C>=5 give R>=25. For z>=2,
distinct classes and blocks give Pi>=14+3z and C>=z. Values z>=3
are excluded. At z=2, R=24 forces Pi=20,A=2,C=2,X=0.
Both Z sources must be the matching edge containing u, with their signed
sum exactly two: otherwise another isolated sum remains nonzero. Thus
rho(u)=-2 and (3) gives C>=4, a contradiction.

We may now assume m=2 or3 and some zeta nonzero.

## 5. Matching: two attachment vertices

Let K=U={u,v}, d count D and z count Z, h=d+z. All D sources are U.
The four inactive vertices and the two active line charges give
Pi>=18+h. At least one matching edge avoids K, so A>=1; if z=0,
all helpers have source U and A>=2.

### 5.1 No Z

Here d>=1. Nonzero flow would cost C>=6 and give R>=19+2+6>24,
so delta=0. For d>=3, (2) gives Pi>=27. For d=2, (2) gives
Pi>=22, forcing Pi=22,A=2,C=X=0. Equality leaves no glue at either
active endpoint, hence Gamma=0, Psi=1/2 and the signed sum of D is
1/2. Then I_U=-2*1_M(U)+1 is nonzero, and at least three isolated sums
are nonzero, contradicting A=2.

For d=1 write its coefficient a. Balance (4) requires NT_U=+1 or -1,
so one endpoint has f>=13; the other has f>=8. Thus
Pi>=ceil(10+13/2+8/2+1)=22. The budget forces A=2,C=X=0 and
Pi=22. The slot law now pins U in M and a=1. Both endpoint Gamma
values are 1/2, so both have f>=9; the only possible split is (13,9).
This order of deductions excludes the otherwise possible (14,8) split.

At the f=9 endpoint all lines belong to its two source-U class lines,
so its separating sum is zero. At the other endpoint the two together
class lines cost four each and a separating line costs five, exhausting
f=13. The cost-four lines have no glue. Therefore its entire Gamma=1/2
lies on the separating line, so NT_U=1/2, contradicting (4)'s value one.
This does not assume extra cancelling N columns are absent; any that fit
the line budget cannot create another glue-bearing line.

### 5.2 At least one Z

Now C>=1. Coverage 2A+L3>=4 implies A+C>=3, hence Pi<=21.
Consequently h<=3, and (2) excludes d>=2. A single Z is impossible
because Pi>=19,A>=1,C>=5. Only (d,z)=(0,2),(0,3),(1,2) remain.

For h=3, Pi=21 exhausts the inactive, active-line and helper charges.
Both endpoints are glue-free: Psi(u)=Psi(v)=1/2 and sum_B z_B=1.
If C>=2, the budget forces A=1. The sole alternating column is on an
unaffected matching edge and has coefficient -2, so (1) instead gives
sum_B z_B=4. If C=1, the two Z must share one block, so the case is
(1,2) with every helper on U. The source sum one leaves I_U nonzero
in addition to the unaffected matching edges, forcing A>=3. Both cases
contradict R=24. This is a signed argument throughout.

For (d,z)=(0,2), Pi is 20 or21.

* At Pi=20 both active line charges are four, so Gamma=0 and there is
  one Z of coefficient 1/2 at each endpoint. All three matching isolated
  sums remain nonzero, so A>=3. The budget forces C=1 and both Z in
  one source-U block. The sole loop must be the double loop; there is
  no L3, hence delta=0. Glue-freeness gives NT_U=0, contradicting (4).
* At Pi=21 the only budgets are A=1,C=2 or A=2,C=1, with X=0.
  In the first, U is not a matching edge, and the two Z are each forced
  to coefficient two on the two matching sources incident to K. Thus
  Gamma(u)=Gamma(v)=3/2. The charge budget forces f=9 at both vertices.
  Their class N coefficients are -2, whose cost-five line sums are two
  or one, never 3/2. In the second budget, both Z share a source-U block.
  Its sole double loop forces equal Z coefficients; the slots force each
  to one and U in M. With no L3 the flow vanishes. Balance requires a
  separating line, so f>=13 at one endpoint and f>=9 at the other.
  This gives Pi>=10+(13+9)/2+2=23, a contradiction.

This exhausts all m=2 helpers, including their boundary budgets.

## 6. Matching: three attachment vertices

Pi>=21. If Pi>=23, coverage 2A+L3>=3 requires at least two other
directions, giving R>=25. At Pi=21 equality forces A>=3, so A=3 and
C=X=0. Some zeta is nonzero, and the equality structure leaves exactly
one D of coefficient 1/2 with glue-free endpoints U; the third active
vertex has only zero-zeta classes. No loops implies delta=0. Equation
(4) still applies with that third vertex present, and contradicts NT_U=0.

At Pi=22 only (A,C,X)=(2,0,0) or (1,1,0) can meet coverage.

In the first case Z=0 and every D source lies inside K. There must be a
matching edge U={u,v} inside K; otherwise three unaffected matching slots
force A>=3. Write w for the third active vertex. The two other matching
edges have isolated sum -2 and exhaust A=2. Hence the signed sums of D
are one on U and zero on every other source. Thus

    Gamma(u)=Gamma(v)=1/2,       Gamma(w)=-1/2.

All three active vertices have f>=9. If d is total D support, disjoint
charges give Pi>=3*(5/2)+3*(9/2)+d=21+d. Since Pi=22, there is
exactly one D, on U, of coefficient one, and f=9 at each active vertex.
The two endpoint class lines are together lines and exhaust their line
budgets, so NT_U=0. Yet C=0 gives delta=0 and (4) requires NT_U=1.
This argument permits signed off-source cancellations and overlapping
private lines; neither can evade the count d.

In the second case coverage forces the sole loop to be L3, not a double
loop. It enters exactly one S row. That row must contain exactly one Z,
and every other Z would violate its own S row. Thus there is exactly one
Z globally, which requires C>=5. This contradicts C=1.

Every matching R=24 case is excluded. Therefore matching retained support
is at least25, completing the global lower28 proof.
