## 1. stated notation and exact identities

Use the canonical model, normalization and notation of
the lower-bound-29 component above:

    R = Pi + A + C + X,       Pi = stars + glues + N + Z + D,
    C = L3 + E,              Pi >= 15 + 2m.

K is the set of N-attachment vertices, m=|K|. Lowercase d,z count D,Z
columns; these are support counts, not sums of their coefficients.
For a block B with source U={u,v}, put

    zeta_u = Z_u+D,  zeta_v = Z_v+D,  beta_B = Z_u+Z_v+2D.

Normalization removes D if either endpoint zeta is zero. Consequently
each surviving D uses a nonzero class at both endpoints; Z can attach
only in K. At a private line let F=g+2s+2n count its glues, stars and N;
f(u) is the sum of F at u. Each private glue is counted twice, so

    2Pi = sum_u f(u) + 2(d+z).

An inactive vertex has f>=5, with first strict increase at least 2.
An active vertex has f>=8, or f>=9 if its total glue marginal Gamma is
nonzero. Gamma(u)=Psi(u)-1/2, with Psi=sum_B zeta_u(B). At f=8 there
are no glues, precisely two N in one nonzero class, and its value is 1/2.
At f=9 with Gamma!=0 there are two N, on a cost5 and a cost4 class line;
all glues at that vertex lie on the cost5 line. The earlier proof's
different-class collision argument is included in this equality statement.

Every vertex has actual star mass1, hence at least one actual star even
when its coefficients have arbitrary signs. Indeed the V-line formula
is S=l-s*1-nu, so its star sum is -2s-sum(nu). Summing at u and
using the two h-side sums gives -2Gamma(u)+2Psi(u)=1. This is consistent
with the stated star-center dual, whose right side is
24*(6*star_center_indicator-star_indicator) before normalizing the
retained coefficients by3. The right side vanishes on all mixed columns.

Each class is (matching group G, source pair U, attachment u in U).
Its four N columns lie on four different private lines, all in halves
containing U. Each of its two h sides sums to -zeta. A nonzero class
needs at least two N; with precisely two, each coefficient is -zeta.

The general array and group identities are

    t_U^G = Omega_G+beta_(G,U),
    sum_(G contains e) Omega_G = -w_e,
    sum_B beta_B = sum_e I_e+6,
    A+L3 >= sum_(Omega_G!=0) #{U in G : U not contained in K},
    2A+L3 >= 6-m.

Every nonzero array total requires its own alternating column. Array
supports are disjoint, including different groups with the same source.
For matching corrections w=2*1_M, the proved signed-incidence lemma
forces Omega=-2*e_M whenever R<=26. Thus the three M-array totals are
beta-2; every other array total is beta.

The projection of the flow and the effective-private flow identity are

    2 sum_B delta_B c_B = sum_B (Z_v-Z_u)(e_u-e_v),
    div(delta)_e = H0_e + sum_B beta_B(1[e=U]+1[e=V]/2+1[e=W]/2)-w_e. (F)

Here c_B=1_W-1_V, U,V,W are the three disjoint pairs of B, and H0 is
computed from effective stars obtained by eliminating N from V. Its
glue-marginal form is H0_uv=(B_uv+B_vu)/2, where

    B_uv = -2 sum_(private halves at u containing v) line_sum(u).

The 15 literal rational certificates for (F) are independently recomputed
by `check_global30_frontier_20261005.py`. A simple inactive vertex has
one line of marginal -1/2, so its B row has two entries 1 and all others
0. A glue-free vertex has a zero B row. H0 need not be nonnegative in
general; the proof only uses these explicitly justified special cases.

At zero flow, summing (F) gives the signed identities

    sum_e w_e=6,                 sum_G Omega_G=-2.                    (M)

To see this without any positivity assumption, each B row sums to
-4Gamma(u), so sum_e H0_e=-2sum_u Gamma(u)=6-2sum_B beta_B.
Each helper contributes beta*(1+1/2+1/2)=2beta to the edge sum.
The terms cancel. Summing the group incidence law then gives
3sum_G Omega_G=-6. All three uses of (M) in section3 have zero flow.

The literal g+g-S-S row combination in every block is

    sum_(four block stars) coefficient
       + sum_(four block alternatings) coefficient = beta_B.           (B)

The 45 block-star supports partition the 180 stars, and their alternating
supports partition the 180 alternating columns. Thus when beta=0 and
actual stars are nonnegative, k blocks with positive star sum force
A>=k, even if the flow is nonzero. This is a separate lower bound on A,
not an addition to the three-array bound. The two counts need not be
disjoint. 
All 45 identities and both partitions are checked against literal rows.

Mixed-column scope is explicit: all180 mixed columns have exactly four
T incidences and no V,h,S,g incidences. They vanish in all six actual
star-center functionals, in all15 general-flow and nu rows, and in the
effective H0/helper-beta rows. The exact h-elimination certificates are
replayed. Thus (F), (B), the center laws and the derived (T) have no
omitted mixed term. This is the P-mix premise required by section7;
it does not extend the fixed21-row dual to mixed columns.

If all nonzero zetas have source U={u,v} and flow vanishes, source balance
gives

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

Zero-zeta classes elsewhere are allowed. A nonzero U-class has two
together lines of cost at least4; a nonzero-sum separating line costs at
least5. Therefore an endpoint with both has f>=13. At f=13 the two
together class lines are glue-free and the separating sum equals Gamma.

Further stated support facts used below are: Z=0 and nonzero flow
imply C>=6; a single Z globally implies C>=5; each block with Z requires
at least one loop column; if all Z lie in one block and C<=4, their two
coefficients are equal and all flow vanishes. A single nonzero loop when
Z=0 is impossible. For a hypothetical matching R=26, the proved
matching lower bound26 makes every support-decreasing kernel move
impossible. In particular

    (D,Zu,Zv,E) -> (D+a,Zu-a,Zv-a,E+a)                                (K)

rules out C=1 and equal nonzero endpoint Z when D or E is present.

Useful charge formulas, including the helper support charge once, are

    m=2: Pi>=18+d+z and Pi>=12+5d+z;
    m=3: Pi>=ceil(19.5+d+z);
    m=4: Pi>=21+d+z.

For Z=0 and D-degree d_u, pi(u)=f(u)/2+d_u/2 is at least
max(4+d_u/2, 1+2.5d_u) when d_u>0. The second bound improves by 1/2
if Gamma(u)!=0. An active vertex with d_u=0 has pi(u)>=4.5.

## 2. Small facts needed at the next budget

**Different prefixes.** At one attachment vertex, classes with different
source pairs have just one private line in common. Its half is the union
of the two pairs. Hence two such nonzero classes use at least three
N-lines; if one has three N and the other two, they use at least four.
The checker verifies all 540 such class pairs. Classes with the same
source but different matching groups are not subject to this statement.

**One Z and one L3 in a block.** The S rows force ell=-Z, while the
literal row identity is 2delta=L_W-L_V. Therefore |delta|=|Z|/2,
and the magnitude multiplying c_B in the projection is |2delta|=|Z|.
Do not conflate delta with its doubled projected coefficient. All 45 row
identities and all 360 compatible Z/L3 placements are checked.
In particular this flow is nonzero. If two single-Z blocks have the same
source and C=2, projection first forces its right side (supported on
that source) to vanish; the two distinct, nonparallel c columns then
force both flows zero, a contradiction.

For clarity, the two S rows of each block have the form
Z_u+E+sum_(four L3 in one set)=0 and
Z_v+E+sum_(four L3 in a disjoint set)=0. D and mixed columns are
absent. Consequently an E-only loop block has Z_u=Z_v=-E (single-loop identity).
A block with opposite nonzero endpoint Z cannot have only one loop:
one E would force equality and one L3 cannot satisfy both nonzero-Z
rows. It therefore needs at least two loops. A Z-free block also cannot
have exactly one nonzero loop, since either E or L3 would violate an
S row. Hence a Z-free nonzero-flow block costs at least two loops.
These direct row consequences justify all loop counts in section8.
Zero flow alone does not imply C=0; whenever C=0 is used, the budget
or the stated S-row argument rules out idle loops as well.

**Low N-lines.** The full line classification, including collisions from
other classes, gives: F4 N-lines are glue-free; any F5 line containing
a specified nonzero N coefficient n has marginal -n or -n/2; an F6
N-line has marginal zero. These are the previously pinned analytic
line controls, not a coefficient grid assumption. If all crossed
inactive vertices have f5, their glues force the grid {0,+/-1/2}.
On this forced grid, a line with its sole N equal to -2 has, for F<=7,
only F4 with sum0 or F7 with sum+/-1/2. All 81 prescribed local forms
are checked. A single nonzero class of value2 has f>=11, as already
proved at lower29.

**Balanced transport.** When all zetas vanish, a balanced h-side pair
(n,-n) changes two effective stars by (-n,+n) on distinct lines of the
same center, preserving V, h, center mass and H. At an active f9 with
effective private support15, n is positive: it cancels either a +1/2
star on a one-glue line or a +1 star on a three-glue line, and its partner
creates the same positive star on an empty line. The checker verifies
all 360 ordered h-side transfers and the nine tight scalar patterns.

**Private support18.** In an effective N=0 private layer whose center
masses are all1, P0<=18 implies H0_e in {0,1/2,1}. The line spectrum is
{5,7,8,9} for a nonzero marginal and {0,6,9} for a zero marginal. There
are at least six nonzero marginals. Eight cost40. Seven cost at least35;
total cost is even, and the next allowed increments are 2,3,4,6,9, so
there is no even cost at most36. Thus there are exactly six, one per
vertex, each sum -1/2; extra zero-marginal lines do not change H0.
The center-mass hypothesis is essential and is used explicitly below.

## 3. Nonmatching and five-edge branches

For a nonmatching correction graph, q=#{G:Omega_G!=0}>=2.
At m>=3 the general charges give R>=25. At m=2, Pi>=20 gives
R>=24. Equality Pi19 has either all zetas zero, giving A>=6, or one
D=1/2, which can cancel at most one of six arrays and gives A>=5.

At m=0, R<=23 forces effective private support15 (the private gap is
15 or at least18), flow zero and A<=8, hence q=2. At m=1, only Pi17
can fit; equality gives all zetas zero, A6,C=X=0. Its effective private
layer has P0<=17, center masses1, hence P0=15. In either case H0=w.
By (M), the two supported Omega values sum to -2. A shared edge
would therefore have weight2, contradicting H0<=1. If disjoint, they
form a six-cycle. Each nonzero weight lies in {1/2,1}, and the two
weights sum to2, so both are1.
Each vertex's halfmates must be its two cycle neighbors. This forces
three single glues and twelve positive stars in twelve different charged
blocks. For m=0, (B) forces A>=12. For m=1, the balanced transport loses
at most one block, so A>=11. Both contradict A<=8 (indeed A=6 at m=1).
The checker verifies all 24 possible balanced transfers of this fixed
six-cycle private core against V, h, H and charged-block incidences.


Now suppose the correction support has exactly five edges. The proved
signed five-edge census gives an edge of weight2. At m>=3, R>=25.
For m=0, A>=6; nonzero flow gives R>=27, while zero flow at R<=24
gives P0<=18, contradicting the preceding H0<=1 lemma. At m=1 a
single Z gives Pi>=18,A>=5,C>=5. Two or more Z give Pi>=20 and
per-group complement>=6. Otherwise all zetas vanish, A>=6, flow is
zero and P0<=Pi<=18 with effective center masses1, again impossible.

For m=2, Pi>=21 already gives R>=25. At Pi20,R<=24 the per-group
charge forces q2 and both supported matchings containing U=K. Their
four outside-U arrays give A4,C=X=0,Z=0. No D would give A6; two D
would give Pi>=22. One D has source U. Since both supported groups
contain U, (M) gives w_U=-(Omega_1+Omega_2)=2, so (T) needs NT=1.
Its separating endpoint f>=13, other endpoint f>=8, inactive cost20
and D charge1 imply Pi>=22, impossible.

At Pi19, all-zero zetas give A>=6. Otherwise equality is one D=1/2
with glue-free endpoints and simple inactive vertices. To fit R24 its
array must cancel one of six; thus q2, A5,C=X=0. Cancellation gives
Omega_D=-1, and (M) then gives Omega_other=-1. Write their matchings
G_D,G'. Formula (F) at zero flow
has beta1 only in the D block. Take the G'-edge at an active endpoint u.
If it is U, (F) requires H0_U=1, but both endpoints are glue-free.
Otherwise it is outside G_D and the helper contribution is zero, so
H0_e=1; one endpoint is glue-free and the other simple, giving H0_e<=1/2.
Contradiction. Thus the five-edge retained support is at least25.

## 4. Matching reductions and all-zero zetas

For the remainder suppose matching R=26, normalized and support-minimal.
Then Omega=-2e_M. If all zetas vanish, Z=D=0 and A>=3. With zero flow
the effective private lemma gives P0>=24, so R>=27. With nonzero flow,
C>=6 and Pi>=17 force Pi17,A3,C6,X0,m1. The effective private layer
has center masses1 and P0<=17, so P0=15. Its positive stars occupy at
least six charged blocks, by the proved private15 classification.
The balanced transport in section2 loses at most one, leaves every
actual star nonnegative, and has beta=0. Identity (B) gives A>=5,
contradiction. N=0 is included using private support15 or at least18:
the first has A>=6,C>=6, and the second has either P>=24 at zero flow
or C>=6. All give R>=27.

Thus at least one helper is nonzero. m6 gives Pi>=27. At m5, coverage
requires another direction; Pi25 equality also forces A>=3, whereas
Pi26 has no budget for coverage. Hence m<=4.

### 4.1 One attachment vertex

D=0. For z>=3, Pi>=14+3z, C>=z, A>=2 give R>=28.

If z=1, its group must be M; otherwise A>=4 and Pi18+C5 already
give27. If A=2 the Z equals2. The single-class bound f>=11 forces
Pi>=19, so equality requires Pi19,C5,A2,X0. All inactive vertices
have f5 and the active one f11. Three N in the nonzero class would
use three lines and cost12. Four N, a star and a glue would exhaust
f11, but the sole glue would have to equal Gamma=3/2, contradicting
the crossed inactive f5 grid. Hence there are precisely two N=-2.
Each of their distinct lines costs at least4; since f11, each costs
at most7. The forced-grid F<=7 check therefore covers this budget.
Their two lines must cost4+4 or4+7, and have total marginal0 or+/-1/2,
not3/2. An additional nonempty glue line cannot fit. Contradiction.

If A>=3, the only remaining budget is Pi18,A3,C5,X0. First, Gamma0
cannot occur at its active f9. Two F4 N-lines leave a cost1 remainder,
which no nonempty line can carry; F4+F5 has nonzero marginal; an extra
balanced class costs at least four N plus one actual star, already10.
Thus Gamma!=0. The active f9 classification forces the Z coefficient1:
Gamma=a-1/2 must equal a
or a/2 from its F5 N-line. Its crossed inactive f5 lines force that
line to be (-1/2,+1/2,+1/2), not a single glue of value1. Its positive
entries force full crossing lines, activating all six vertices in one
private split. The marginal totals of its two halves would then be
1/2-1/2-1/2=-1/2 and -1/2-1/2-1/2=-3/2. They must be equal, a
contradiction. 

Let z=2. Pi>=20,C>=2,A>=3; at most one Z block belongs to M.
The possible budgets are Pi20,A3 (with C2 or3), Pi20,A4,C2,
or Pi21,A3,C2; X uses any otherwise unspent unit.
In the first A3 case the M coefficient is2. Write b for the other.
The active f11 and inactive f5 force exactly four N, one actual star
and one glue. Its coefficient is +/-1/2, so Gamma=2+b-1/2 makes
b=-2 or-1. Every N is therefore an integer. A single +/-1/2 glue
creates two nonzero effective half-star entries, which integer N cannot
both remove; one actual star is impossible. This argument does not
depend on C or X.

For Pi20,A4,C2, both loops are single L3s, one in each Z block.
Projection at the attachment vertex gives a+b=0. Different source
pairs would require three N-lines, costing12>f11. Equal source pairs
instead give a zero projection and two nonparallel c columns, forcing
zero flows contrary to section2. This excludes R1c.

For Pi21,A3,C2, the coefficients are2 and -2 by the same projection.
Equal source pairs are again impossible. With different pairs, four
N use at least three lines. Five N would use at least four lines,
and six N plus a star and a glue cost15. Thus there are exactly four
N, all +/-2. The active f is at least13; since total cost is38, it
is exactly13 and every inactive vertex has f5. Its three lines cost
5+4+4. The only glue-bearing line contains an N of magnitude2, so
its sum is +/-2 or +/-1, not Gamma=-1/2. This excludes R1b and
completes m1.

## 5. Four attachment vertices

Pi>=23. At Pi23 equality permits no Z and only disjoint D=1/2,
with glue-free endpoints. Equality needs A>=3, hence A3,C=X=0.
Every D block must belong to M, since an outside-M array would add A.
One D contradicts (T): its endpoints are glue-free. For two D on
disjoint M-edges U,V, formula (F) requires H0_U=1/2; its endpoints
are glue-free, so H0_U=0. This excludes R4a.

At Pi24 with Z present, C>=2 leaves A0,C2,X0. To cancel all three
M arrays, K must contain one whole M-edge U and one endpoint of each
other M-edge. The two crossing M-blocks need Z=2. Since d+z<=3,
the remaining helper must be D=1 on U; a third Z block would require
C>=3. All four active Gamma values are nonzero, so f>=9 at each
and Pi>=5+18+3=26, contradiction.

Now Z=0. A0 is impossible because not every M-edge is inside K.
At Pi24 this leaves C<=1, and the S rows exclude a lone nonzero loop,
so C=0. A1 forces D=1 on two M-blocks whose
sources cover K. Their four Gamma values are1/2, giving Pi>=25.
Thus only A2,X0 remains at Pi24. At least one M array must cancel,
requiring D=1 on its block (M,U).

Three or more D already cost Pi>=26: the least D-degree distribution
on four active vertices is (2,2,1,1), which costs6+6+4.5+4.5 plus5
for the inactive vertices. More concentrated distributions cost more.
With one D, all four active vertices have Gamma!=0. Pi24 forces
every active f9, so its two D endpoints have only together glues and
NT_U=0, contrary to (T). With two D, an outside-M block would force
an extra array; cancelling two M arrays as well would need two M D
columns, exhausting d2. Therefore the second D is on another M-edge
V, with coefficient a. If a!=1/2 all four Gamma values are nonzero,
giving Pi>=25. If a=1/2, Pi24 forces f9 at U, f8 at V, and inactive
f5. At U all marginal1/2 lies on together lines, so H0_U=-1.
Formula (F), with betas2 and1 on U,V, instead requires H0_U=-1/2.
This excludes R4b.

At Pi25, A1,C=X=0 is forced. There are precisely two D=1 on disjoint
M-edges U,V. All active f9, inactive f5. At each D endpoint the F5
line has N=-1, sum1/2, three glues (-1/2,+1/2,+1/2), and no actual
star. Its F4 line gives a positive unit star. Every actual star is
positive. The F4 stars at the two endpoints of U have different source
pairs, neither U nor V: each source contains the other endpoint of U
and a third vertex. Their charged blocks therefore have beta0 and
are distinct. Identity (B) forces A>=2. The 720 literal source-incidence
controls verify this claim. This excludes R4c. Pi26 leaves no coverage
budget. All m4 cases are excluded.

## 6. Three attachment vertices

First Z=0. Nonzero flow cannot fit; d0 was excluded in section4.
For d1, (T) forces one endpoint f>=13. The other endpoint is at least8,
the third active vertex has f>=9 and the inactive cost is15, so Pi>=24.
To fit R26 requires A2, D=1 on an M-edge U, and C=X=0. Both D
endpoints then have Gamma1/2. The tight costs are (13,9,9) at active
vertices and (5,5,5) inactive. The separating sum is1/2, not NT_U=1.

For d2 with a common source, the charges6+6+4.5+7.5 give Pi>=24;
at least one outside-M array makes A>=3. For different sources, they
form a path on K and Pi>=23,A>=3. Equality R26 requires one M D=1
on U and a second, outside-M D of coefficient a. The common endpoint
has charge at least6+(1/2)[a+1/2!=0], the M-leaf at least5, and the
other leaf at least4.5+(1/2)[a-1/2!=0]. Add inactive7.5. The two
indicators cannot both vanish, so Pi>=24, contradiction. For d>=3,
the degree count gives Pi>=26, before the at least two M arrays.

For Z present, a single Z has C>=5 and Pi>=22, impossible. Thus
z>=2,C>=2. At Pi22, the formula Pi>=ceil(19.5+d+z) forces d0,z2.
The total line cost40 is just one above the minimum39. Hence the two
Z are1/2 at different vertices, with f8 there and f9 at the remaining
active vertex. No M array cancels, so A>=3 and R>=27.

At Pi23, put h=d+z<=3. If A0, K must meet all three M-edges once,
and cancellation requires three Z=2 in three blocks. Each vertex then
has f>=11, giving Pi>=27. The remaining possibility is A1,C2,X0.
If K contains an M-edge U and a third vertex w, an untouched M-edge
uses the one A. Thus both the U array and the crossing M-array at w
must cancel. The latter needs Z_w=2. For h2, the other Z must be2
at an endpoint of U; f>=11,11,9 yields Pi>=25. For h3, all remaining
helpers must be in the U block: either D plus one Z, or two Z.
The costs f(w)>=11 and f>=8 at the other two vertices already give
Pi>=24. A helper elsewhere would create an additional array.

If K meets all three M-edges once, A1 requires Z=2 at two vertices.
For h2 their two f>=11 and the remaining f>=9 give Pi>=25. For
h3, a third Z in the remaining M-block gives Pi>=26; a third Z
elsewhere or a D creates an extra outside-M array while the remaining
M-array still needs A. At Pi>=24, A0,C2 is the only possible residual
budget, but canceling all M-arrays either is impossible because an
M-edge avoids K or needs three separate Z blocks and C>=3. This
completes m3 by the stated cases.

## 7. Two attachment vertices, no Z

Let U=K. The flow is zero: otherwise Z=0 requires C>=6, and even the
minimum Pi19 and two unavoidable M alternatings give R>=27.
Three D cost Pi>=27. With two D, each endpoint has at least
four N and a star, hence f>=10. Source balance requires a separating
line at an endpoint, so their f-sum is at least13+10. Add inactive20
and two D: Pi>=24. Their distinct blocks give A>=3, impossible.
Thus there is one D, coefficient a, and flow is zero.

Source balance gives Pi>=22. In fact Pi22 is impossible. If a!=1/2,
both Gamma values are nonzero and the only endpoint costs are13,9,
with inactive f5. At f9, the F5 marginal a or a/2 must equal a-1/2,
forcing a=1. At f13 the separating marginal equals Gamma=1/2,
whereas (T) requires +1 or-1.

If a=1/2, Gamma=0 at both endpoints. A separating endpoint has f>=13,
but equality13 makes its together lines F4 and its separating marginal
equal Gamma0, impossible. Hence f>=14. Total cost42, inactive cost20
and the other active cost8 force exactly (f_u,f_v)=(14,8), with every
inactive f5. The f8 endpoint is glue-free, so the f14 endpoint alone
has separating sum S=NT=+/-1 and together sum T=-S. Its main nonzero
U-class needs two distinct together lines, each cost at least4. Three
main N would use three such lines, costing12 before a separator of
cost at least5; hence there are precisely two main N=-1/2. Any
additional together line, including a balanced-N-only line, costs at
least4 and would give at least4+4+4+5=17>14. Thus all together
marginal is in the two main lines. Their combined cost is at most9,
giving (4,4) or(4,5). The full collision-safe F4/F5 classification
gives T in {0,1/4,1/2}, even with additional balanced N colliding
with a main N. This contradicts T=+/-1. No arbitrary balanced-class
coefficient has been bounded or enumerated. Therefore Pi>=23.

A D outside the group-M block of an M-edge would leave all three M
arrays and add its own, so A>=4 and R>=27. Therefore D is on (M,U)
with U in M. The two other M-arrays require A>=2. No outside-M array
can carry an alternating column when A<=3: its zero total would need
two such columns in addition to these two. C=0. With X0 the stated
21-row dual on all private columns, all N at U, all M alternatings and
this D reads -a+t_U^M/2=0. Since t_U^M=-2+2a it gives -1=0.

The only remaining mixed budget is Pi23,A2,C0,X1,a=1. Both endpoint
Gamma values are1/2, NT_U=1, and their f-sum is at most24. If a
separating endpoint has f13, its separating marginal is1/2, requiring
a separating line at the other endpoint as well; then both f>=13,
impossible. Otherwise its f is14 or15 and the other is at most10,
so the first alone has separating marginal1. Its together marginal
must be -1/2. Its two main U-class lines cost at most10, since a
separating line costs at least5. A third main-class N would already
use three lines, costing12 before the separator. Thus each of the two
main N coefficients is -1. Additional N from balanced classes may
collide in these lines; the full F5/F6 classification allows that.
The possible two-line costs are (4,4),(4,5),(5,5),(4,6); each has
nonnegative total marginal. An additional together line, including a
balanced-N-only line, would cost at least4, giving at least8+4+5=17>15.
The required together
marginal -1/2 is impossible. 

## 8. Two attachment vertices, Z present

There is an M-edge avoiding K, so A>=1. Minimality gives C>=2 and
coverage gives A+C>=3; consequently Pi<=23. One Z would have
Pi>=20,C>=5,A>=2 and cannot fit. Thus z>=2. The charge formulas
force d<=1 and h=d+z<=5.

For h5, Pi23 is equality in Pi>=18+h: both active f8, inactive f5.
The active Gamma values vanish, so sum beta=sum Psi=1. The budget
forces A1,C2,X0. Its sole alternating is on the unaffected M-edge
with coefficient -2, so the total law instead gives sum beta=4.
Contradiction.

For h4, equality Pi22 would give one nonzero class per endpoint.
Four Z without D use too many classes; one D fixes the classes at
both endpoints, but only two Z can share that block, not three.
At Pi23 the budget is A1,C2,X0. K must meet two M-edges, and their
two arrays must cancel by Z=2 in their separate M-blocks. The other
two helpers must cancel their beta outside M. They are either a pair
of opposite Z or a D plus Z=-2D in a third block. Either placement
requires a third Z block and C>=3. Contradiction.

For h3, Pi21 again has active f8, hence one nonzero class per endpoint.
Three Z are impossible. One D and two Z must share one block, with
equal Z coefficients1/2-D; move (K) lowers support, impossible.
A1 is impossible for h3 at any Pi: it would need two Z=2 in the
two affected M-blocks and a third helper outside them, giving an extra
array. Thus the only h3 budget is Pi22,A2,C2,X0.

If U belongs to M, the two other M-arrays exhaust A and beta_(M,U)=2.
For d0,z3, one Z=2 in that block leaves an opposite Z pair in another
block; that pair requires at least two loops, in addition to the first
block's loop. Two Z in the M-block would leave an uncancelled third
Z elsewhere. For d1,z2, putting all helpers in the M-block invokes
the one-block lemma and equal-Z kernel move. D=1 in the M-block and
opposite Z elsewhere also violates the one-block lemma. The only
other zero-array placement is an outside-M D canceled by Z=-2D,
and Z=2 in the M-block. These are two single-Z blocks with the same
source U and only two loops, excluded by section2. Other placements
leave a nonzero outside-M array.

If U is outside M, its unaffected M-edge consumes one A. For d0,z3,
canceling both affected M-arrays needs two Z=2 and leaves a third
Z block, so C>=3. Canceling just one leaves two opposite Z in a
block outside M; they require two loops plus the first block's loop.
Canceling neither needs three M alternatings. For d1,z2, canceling
both affected M-arrays uses Z=2 at each endpoint and leaves a D array.
Each endpoint then has two different nonzero classes, at least four
N and a star, so f>=10 there and Pi>=10+10+3=23, contradiction.
Canceling neither crossing M-array gives A>=3. Canceling just one
requires a D/Z block with beta0 outside M. At the
endpoint of the M Z, its source and the D source differ, so three
N-lines give f>=12; the other endpoint has f>=8. Then Pi>=23 again.

It remains d0,z2, Pi>=20. If both Z lie in the same block, A>=2 and
C<=4. The one-block lemma gives equal Z and zero flow. Source
balance needs a separating endpoint f>=13, with the other at least8;
including inactive20 and z2 gives Pi>=23. With A>=2,C>=2 this is27.

If both distinct Z blocks are outside M, A>=5 gives R>=27. If both
are in M, the projection's right side has opposite nonzero values on
the affected M-edge endpoints, while every M-block c is constant on
each M-edge. A non-M flow block is therefore necessary, costing two
loops in addition to the two Z blocks: C>=4. With A>=3 this is27.
For A2 at least one Z is2, giving Pi>=22; for A1 both are2, giving
Pi>=23. These are even larger totals.

Finally suppose one Z block is in M and one is outside. A>=3. At
A3 the M Z equals2; if they attach at different vertices, f>=11+8;
if at the same vertex, four N and a star cost at least10 there and
the other active zero-zeta vertex costs9. Either way Pi>=22, so
R>=27. The only remaining budget is Pi20,A4,C2,X0, with one Z=1/2
at each glue-free endpoint, inactive f5 and one L3 in each Z block.
Let e be the source of the M Z. At its active endpoint B=0; its other
endpoint is either also active or simple inactive. Hence H0_e<=1/2.
The first helper contributes1/2 to (F). The second contributes1/2 if
e is its source, 1/4 if e is its side, or0 otherwise. If e is its
source, both flows have zero divergence there, while (F) is at most
1/2+1-2=-1/2. If it is a side, the required divergence is at most
1/2+3/4-2=-3/4, while the sole possible contributing flow has magnitude
1/4. If neither, the required divergence is at most-1 and the actual
divergence is zero. Contradiction in every case. This excludes R2c,
R2d and R2e, including both placements U in M and U outside M.
The checker verifies all 112 prescribed rational corner controls.

Thus every matching identity has retained R>=27.
