## 1. Established inputs and notation

Use the notation, normalization, exact identities and support counts of
the lower-bound-28 component above. Thus

    R=Pi+A+C+X,     Pi=stars+glues+N+Z+D,     C=L3+E,
    Pi>=15+2m,     Gamma(u)-Psi(u)=-1/2,
    t_U^G=Omega_G+z_(G,U),       sum_(G contains e) Omega_G=-w_e,
    sum_B z_B=sum_e I_e+6,
    A+L3 >= sum_(Omega_G!=0) #{U in G: U not contained in K}.

K is the set of N-attachment vertices; m=|K|. In the matching case,
w=2*1_M. A is alternating support, C loop support, X mixed support.
For a private line, F=glues+2*stars+2*N; f(u) is its sum at u.
An inactive vertex has f>=5. An active vertex has f>=8, or f>=9 if
Gamma is nonzero. Equality Pi=15+2m permits no Z and only disjoint D
columns of coefficient1/2, with glue-free endpoints; in the matching
case it forces A>=3. All-zeta-zero matching identities with N>0 have
R>=26: otherwise effective private support is below24, hence nonzero
flow costs C>=6 and Pi>=17,A>=3.

The normalization is support-nonincreasing and preserves the correction
graph and N. Since matching R>=25 is already proved, any hypothetical
R=25 identity, before or after normalization, admits no support-decreasing
exact kernel move. This minimality is used only for the explicit
four-column relation

    (D,Zu,Zv,E) -> (D+a,Zu-a,Zv-a,E+a).                 (1)

Whenever Z=0, nonzero flow costs C>=6. A single nonzero loop column with
Z=0 is impossible by its S rows. A single Z globally costs C>=5.
Every block containing Z needs a loop column. Source-pair balance says
that if every nonzero zeta has source U={u,v} and flow vanishes, then

    NT_U := sum_(private splits separating u,v)(s(u)+s(v)) = w_U-1. (2)

If an endpoint has a nonzero U-class and a nonzero-sum separating line,
its class lines cost4+4 and that extra line costs5, so f>=13. Zero-zeta
classes at other vertices are permitted in (2).

All these inputs have literal rational row certificates in the lower28
checkpoint. The proof also uses the private P15 incidence checks and the
extended fixed21-row dual already pinned there. Neither the two-private-
partition theorem nor the six-star/all-correction theorem is needed.

## 2. A signed kernel lemma forces one matching group

**Lemma.** Any nonzero linear dependence among distinct perfect-matching
incidence vectors on six vertices has at least six nonzero coefficients.

Distinct perfect matchings share at most one edge. In a dependence, every
edge of a supported matching must occur in another supported matching.
Thus every matching needs at least three other matchings, one for each
edge, and support below four is impossible.

With four matchings, an edge shared by three would leave a containing
matching's other two edges to be covered by the sole remaining matching,
which cannot share two edges with it. Thus no triple-shared edge exists.
Every pair must intersect, each intersection edge occurs exactly twice,
and its coefficients are negatives of one another.
A triangle in this intersection graph makes every coefficient zero.

With five matchings, first suppose no edge occurs three times. Then every
matching has exactly three neighbors in the intersection graph. This
would be a 3-regular graph on five vertices, impossible by parity.
An edge cannot occur more than three times in K6. If three matchings
G1,G2,G3 share an edge, their other two edges must each be covered by the
two remaining matchings H1,H2, one edge from each. The six other edges
of G1,G2,G3 are distinct. Each Hi therefore contains one from each Gj.
Their edge equations force all three Gj coefficients to be equal and
opposite to either Hi coefficient. The shared-edge equation then reads
3a=0, a contradiction. This proves the lemma over Q (indeed over R).

For matching corrections, Omega+2*e_M is in this kernel. Unless it is
zero, the support q of Omega is at least five. Every matching has at
most floor(m/2) edges inside K. The per-group charge would therefore give

    R >= 15+2m + 5*(3-floor(m/2)).

For m=0,...,6 these bounds are respectively30,32,29,31,28,30,27.
Consequently

    matching R<=26  implies  Omega_M=-2, all other Omega_G=0.  (3)

For the rest of the matching proof assume R=25 and use (3). Each array
in M has total z_B-2; each array outside M has total z_B. Every nonzero
array total requires its own alternating column. These arrays are
disjoint, including arrays with the same source edge in different groups.

## 3. Additional small local facts

**Two helper counts.** If Z=0 and vertex u meets d_u different D blocks,
its h classes require at least2*d_u N columns, and its centered star mass
requires at least one star. Hence

    pi(u) >= 1+(5/2)*d_u.                                (4)

For m=2 and d D columns, this gives Pi>=12+5d+z even when Z is present:
normalization makes both zetas of every D block nonzero. Z is an additional
support charge; it is not assumed to require additional N when it shares
a D block. Also for m=2, Pi>=18+d+z. For m=3, Pi>=ceil(19.5+d+z).

**Equal Z coefficients.** If Zu=Zv=a!=0 in a block with D!=0, move (1)
deletes both Z and leaves at most D and E, saving at least one direction.
If that block has D=0 and its only loop column is E, the S rows give
E=-a, and (1) replaces its three Z,Z,E terms by one D. Thus an R=25
matching identity with Z present cannot have C=1. All its Z would lie
in one block. A sole L3 has only one nonzero S incidence, so the other
S row forces one endpoint Z to vanish; exactly one Z then remains
and the single-Z bound requires C>=5. A sole E instead forces equal
endpoint Z and the support-decreasing move just described.

**Z in one block.** Suppose all Z are in one block B with source {u,v},
and C<=4. The coordinate projection of the flow has right side
(Zv-Zu)*(e_u-e_v). The column c_B is supported on the other four
coordinates. If Zu!=Zv, after subtracting B's flow the residual has
support two or six. A single other flow column has support four. At least
two Z-free flow blocks are necessary, costing four loops in addition to
B's one, a contradiction. Hence Zu=Zv. The projection is now zero;
nonzero flow needs at least three distinct blocks and again costs at
least1+2+2=5. Therefore all flow vanishes. This argument also allows B's
flow to be zero and arbitrary signs.

**The one-N line of cost six.** Let its unique N coefficient be n!=0,
in position i. For F=6, the only possible support counts are two glues
and one star: zero glues would produce exactly one star and F=4.
Solving `S=l-s*1-n*e_i` with those supports forces, up to the other
position j, `l=n*e_i-n*e_j`, `s=0`, and `S=-n*e_j`.
Thus the line sum is zero. The existing F=5 forms have sum -n or -n/2;
F=4 has no glue.

In particular, suppose an active vertex has exactly one nonzero zeta
class, of value a=2, and no other helper there. Then Gamma=3/2 and
**f>=11**. If f<=10, there are exactly two N columns: three in a nonzero
class use three lines and cost at least12; four N plus a star and a glue
cost at least11. Each of the two N coefficients is -2. The only two-line
costs up to10 are (4,4),(4,5),(4,6),(5,5), with total glue sums in
{0,1,2,3,4}, never3/2. A third nonempty line would cost at least5 or6.

For later equality cases, f=8 and no glue forces two cost4 N-lines.
Each has either one star and one N, or no star and two N in the same
V position with opposite coefficients. Two star-free lines contradict
center mass one. One line of each kind would give three N. The two
cancelling columns in the star-free line belong to different prefix
classes, as the literal V incidences verify. Both classes would need
partners elsewhere; the third N cannot supply both. Thus at least four
N would be needed, contradicting the cost. In the remaining case each
line has one N and one star. They cannot be a zero-zeta cancelling pair
because their star sum would vanish; they belong to one nonzero class
and center mass one makes that class value1/2. At f=9
with Gamma!=0 there are exactly two N columns on cost5 and cost4 lines;
their only glue-bearing line belongs to that class.

## 4. Matching: N=0, all-zeta-zero, and m>=4

For N=0, zero flow requires private support P>=24 and A>=3. Nonzero
flow costs C>=6. R=25 would then give P<=16. The signed private line
spectrum is {5,7,8,9} for a nonzero marginal and {0,6,9} for a zero
marginal. P<=16 leaves exactly six nonzero marginal lines, each sum-1/2,
and no nonempty zero-marginal line. P=16 would require one cost7 line.
Its two entries outside {0,+/-1/2} force two other crossing lines to have
cost above5, impossible. Thus P=15.

At P=15, every star is positive and every line has cost5. If there is a
three-glue line, its positive entry forces a full crossing line, so all
six vertices use that same split. Its stars occupy distinct loop blocks,
and there are at least six. Otherwise all lines are single-glue lines:
three glue generators give twelve positive stars and no block receives
more than two. In either case A>=6 by the g+g-S-S block identities. This
incidence statement is already checked on all ten splits and the216
prescribed generator triples. Hence R>=15+6+6=27, contradiction. This
proves the needed N=0 exclusion without the stronger partition theorems.

All-zeta-zero N>0 was excluded in section1. Suppose some zeta is nonzero.
For m>=5, Pi>=25 and coverage requires at least one additional direction
when m=5; m=6 gives Pi>=27. At m=4, Pi>=23. Equality forces A>=3.
Pi>=25 is impossible by 2A+L3>=2. At Pi=24 the sole remaining direction
must be A=1,C=X=0: A=0 would require at least two L3 directions. The
single alternating column leaves two zero M arrays. Thus Z=0. Two M arrays must be zero, requiring D=1
in each of their two blocks, whose four endpoints are exactly K; no other
D block can have a nonzero array. Each active Gamma is1/2 and has f>=9.
This gives Pi>=2*2.5+4*4.5+2=25, contradiction.

## 5. Matching: one attachment vertex

D=0. Let z count Z columns. If z>=3, their disjoint h classes give
Pi>=14+3z, C>=z, A>=2, already R>=28. Only z=1 or2 remains.

For z=1, Pi>=18,C>=5,A>=2 force Pi=18,C=5,A=2,X=0.
The two unaffected matching slots exhaust A, so the Z source is the
matching edge at u and its coefficient is2. All five inactive vertices
have f=5 and u has f=9. But its Gamma=3/2 contradicts section3's f>=11.

For z=2, Pi>=20 and C>=2. At most one Z block can be a group-M block;
the other has a nonzero array outside M. The two unaffected M arrays
therefore give A>=3. Equality R=25 forces Pi=20,A=3,C=2,X=0.
The unique group-M Z is2, cancelling its array; write b for the other Z.
At u there are at least four N columns and one star, so f(u)>=10.
The five inactive vertices each cost at least5, and their first possible
excess is2. Since sum f=2*(Pi-z)=36, they all have f=5 and f(u)=11.
Thus u has exactly four N, one star and one glue. That glue crosses an
inactive cost5 line, so its coefficient is +/-1/2. Therefore
2+b=Gamma(u)+1/2 is0 or1, giving b=-2 or-1.
Every N coefficient is now an integer: each class has just one column
on each side. The single glue creates two effective-star entries of
value +/-1/2. Subtracting integer N contributions leaves both nonzero,
contradicting the single actual star. This excludes m=1.

## 6. Matching: three attachment vertices

First let Z=0 and d count D. The case d=0 is the previously excluded
all-zeta-zero case. Nonzero flow cannot fit R=25, since
Pi>=21,A>=2,C>=6; hence flow is zero when needed below.

For d=1, balance (2) forces a separating line at an endpoint of its
source U. Its f is at least13, the other endpoint at least8, the third
active vertex at least9, and the inactive sum at least15. Including D,
Pi>=ceil((13+8+9+15)/2+1)=24. At least two M arrays remain, so R>=26.
For d=2 with a common source, the endpoints each have charge at least6
by (4); the third active vertex costs4.5 and the inactive vertices7.5.
Thus Pi>=24 and A>=2. If the two sources differ, their shared endpoint
costs6 and the two other endpoints at least4.5 each, giving Pi>=23.
At most one D can cancel an M array, while an off-M D supplies an extra
nonzero array; thus A>=3. With d>=3, (4) summed over K gives
Pi>=ceil(7.5+3+5d)>=26. All Z=0 cases are excluded.

Now let Z>0. Pi>=22, because equality at21 has no Z. A single Z costs
C>=5 and is impossible. Write h=d+z; the bound
Pi>=ceil(19.5+h) shows that at Pi=22 there are exactly two Z, no D.
The total line budget has just one unit above three inactive costs5 and
three active costs8. At least two active Gamma values must be zero.
The two Z must therefore attach at different vertices and both equal1/2;
the third vertex has Gamma=-1/2. No M array can be cancelled by this
total helper mass1, so A>=3. With C>=1 this gives R>=26.

If Pi>=23, at most two directions remain. C=1 is ruled out by the kernel
move in section3. The only possible budget is A=0,C=2. If K contains an
M edge, another M edge avoids K and forces an alternating direction.
Otherwise K contains one endpoint of each M edge. To cancel all three
M arrays then requires a Z in each of their three separate blocks, hence
C>=3. This excludes m=3, including all mixed-column budgets.

## 7. Matching: two attachment vertices, no Z

Write K=U, d=#D. Each D has source U. Pi>=12+5d excludes d>=3.
If d=2, Pi>=22. At least one of the D blocks is outside M, or U itself
is not an M edge. Formula (3) therefore gives A>=3. R=25 forces
Pi=22,A=3,C=X=0. Equality in the count12+5d leaves no glue at either
active endpoint. Balance (2) instead requires NT_U=+1 or-1. Contradiction.

Let d=1. Nonzero flow costs C>=6 and cannot fit; balance (2) forces
f>=13 at one endpoint, so Pi>=22. If the D block is outside M, all
three M arrays and its own extra array are nonzero, giving A>=4 and
R>=26. This also excludes U outside M, since no M block then has source U.
Thus U belongs to M and the D is in the block (M,U), with coefficient a. At most three
alternating columns remain. No outside-M array can be active: it would
have zero total and need two columns, in addition to the two unchanged
M arrays. There can be no loop column: C<=1 and Z=0, whereas every L3
has a nonzero S incidence and E has two. A sole nonzero loop cannot
satisfy those homogeneous S rows. Therefore C=0.

If X=0, the previously checked21-row dual on all private columns, N
attached anywhere in U, all group-M alternatings and this D reads

    -a+t_U^M/2=0.

But t_U^M=-2+2a, making its left side -1. If X=1, the budget forces
Pi=22,A=2,a=1. Both Gamma values are1/2 and both f>=9. Balance forces
a separating line, so the only charges are (13,9), with inactive f=5.
The cost13 endpoint has two glue-free cost4 together lines and a cost5
separating line. Its separating sum equals its entire Gamma=1/2. The
cost9 endpoint has only together lines. Thus NT_U=1/2, contrary to (2)'s
value1. This also excludes the one-mixed-column boundary.

## 8. Matching: two attachment vertices, Z present

At least one M edge avoids K, so A>=1. Coverage 2A+L3>=4 and C>=1
give A+C>=3, hence Pi<=22. A single Z has Pi>=20 and C>=5, already
R>=26. Thus z>=2. Since Pi>=18+d+z and Pi>=12+5d+z, only d<=1
and h=d+z<=4 remain. C=1 is excluded by section3; hence C>=2.

**h=4.** Pi=22 and all inactive f=5, active f=8. The active Gamma
values vanish, giving sum z_B=1. The remaining budget would require
A=1,C=2. This implies U is outside M; otherwise two M edges avoid K
and require two alternating columns. Its sole alternating column is on
the unaffected M edge and
has coefficient-2, so sum I=-2 and the exact total law instead gives
sum z_B=4. Contradiction.

**h=3, Pi=21.** The same equality forces active f=8, no glue and one
nonzero class of value1/2 at each endpoint. Three Z without D would
occupy too many classes. With one D and two Z, all must share one block,
and the Z coefficients are equal to1/2-D. Move (1) saves at least one
direction, contradicting the proved matching lower bound.

**h=3, Pi=22.** The remaining budget is A=1,C=2. K must meet two
different M edges, and each corresponding M array must be cancelled.
This needs a Z of coefficient2 in each of those two blocks, since D
has source K, outside M. Their other endpoints are inactive, so no
additional Z can share either of those blocks. With (d,z)=(0,3), the
third Z therefore lies in another block and gives an extra nonzero array. With (1,2), the D gives that
extra array. Neither can fit A=1, which is already on the unaffected
M edge. Thus h=3 is excluded.

It remains that **d=0,z=2**, with 20<=Pi<=22. For Pi=20,21,22 the
respective upper bounds on A are3,2,1. The two Z may share a block.

If their blocks differ, each outside-M Z block gives an extra nonzero
array. With one such block A>=3; with two A>=5. If both are in M,
their attachment vertices are on different M edges. A<=2 forces at
least one coefficient to equal2, and A=1 forces both to equal2. The
value2 endpoint has f>=11 by section3, while the other endpoint has
f>=8. Hence Pi>=ceil(10+2+(11+8)/2)=22. This excludes A<=2 at
Pi<=21; A=1 at Pi=22 forces both endpoint costs>=11, hence Pi>=23.
The same argument excludes the one-M/one-outside-M configuration at
A=3 and Pi=20, since its M coefficient must equal2 to cancel that array.
If both Z attach to the same vertex, their four N plus the other active
vertex's cost9 already force Pi>=22, and their distinct blocks give
A>=3. Thus no case was lost by the endpoint bounds above.

If both Z share a block outside M, the three M arrays remain nonzero.
If they share a block in M, A<=2 requires their sum to equal2.
All Z are then in one block and the remaining loop budget is at most3;
section3 gives equal coefficients and zero flow. Thus both equal1.
Both endpoint costs are at least9, forcing Pi>=21. At Pi=21 they both
have f=9, only together class lines, so NT_U=0; (2) requires1.

The only budget still to consider is Pi=20,A=3,C=2,X=0. Equality fixes
inactive f=5, active f=8, with one Z of coefficient1/2 at each endpoint.
If they share an M block, the one-block lemma makes the flow zero,
and glue-freeness contradicts (2). A shared outside-M block would need
A>=4. For distinct blocks, one outside M would need A>=4 because its
group-M companion has coefficient1/2, not2; two outside M need A>=5.
Thus the only remaining placement is in two distinct group-M blocks.
The two loops must lie one in each Z block. Their projected columns
are constant on both coordinates of every M edge. The right side of
the projection identity is instead opposite nonzero values on each
of the two affected M edges (one endpoint has Z=1/2, the other zero).
This is impossible. All m=2 cases are excluded.

We have excluded every matching identity with R=25. The earlier bound
excluded smaller R, so matching retained support is at least26.

## 9. Six corrections cannot have retained support22

Six correction edges require at least two nonzero Omega groups, q>=2.
For N>0 the per-group charge gives, respectively at m=1,2,3,

    R>=17+6=23,       R>=19+4=23,       R>=21+4=25.

At m>=4, Pi>=23 alone suffices. Thus R=22 would force N=0 and
normalized Z=D=0. Then A>=3q>=6, so P<=16. Nonzero flow would cost
C>=6 and give R>=27; hence flow is zero and H=w.

The private line count makes every vertex simple and every nonempty
zero-marginal line impossible: its cost6 exceeds the slack above six
cost5 lines. Therefore H entries are in {0,1/2,1}. Also A<=7 implies
q=2. Two perfect matchings supplying six correction edges must be
edge-disjoint; their union is a six-cycle. The correction degree2 at
each vertex and H<=1 force weight1 on every cycle edge.

The directed private formula H_uv=(B_uv+B_vu)/2 now forces each simple
vertex's two halfmates to be its cycle neighbors. Opposite vertices
are the only active centers in each of three complementary3|3 splits.
Since zero-marginal lines vanish, each such private matrix has a single
glue of coefficient-1/2 and four positive stars. The six center-halfmate
source pairs are distinct, and each center's two stars occupy distinct
blocks. Thus all twelve stars occupy twelve distinct charged blocks,
each requiring an alternating direction. This contradicts A<=7.

This last prescribed six-cycle incidence is checked directly against the
literal V equations, H rows, and all star-bearing block rows. No six-star
subclass assumption enters it: this private core has twelve stars.
