Use the lower bound 31 notation and equations. K={u,v}, h=d+z,
Pi>=19, Pi>=18+h, R=Pi+A+C+X, 2Pi=sum f+2h. At least one M-edge
misses K, so A>=1. Inactive f>=5 with no f6; active f>=8, or9 if
Gamma!=0. A vertex with two nonzero classes costs at least13. Actual
star mass is1 at every center. Every surviving D has source K.

## 1. Local refinements and projection controls

### 1.1 Two classes at cost13 have equal absolute values

Suppose a center has two nonzero classes a,b and f=13. Then |a|=|b|.
Six N plus an actual star already cost14. Five main N occupy at least
four lines for distinct prefixes; for the same prefix the proved
different-position incidence makes their minimum cost5+5+4=14.
Thus exactly four N remain, two per class, and no balanced class fits.

For three occupied lines, their costs are4,4,5. The shared two-N line
must have F4 or F5. F4 requires opposite coefficients at the same V
position, giving a=-b; F5 requires equal coefficients at distinct V
positions, giving a=b. Two occupied lines can occur only for equal
prefixes, with different V positions on both lines. Each costs at
least5. At total13, at least one costs5 or6; the former requires a=b,
the latter a=-b. Four occupied lines cost at least16. This proves the
claim, including the same-prefix case.

### 1.2 Values2,2 or2,-2 cannot occur at f13

In the three-line case, all marginal lies in an F5 line and belongs
to{+/-1,+/-2}. It cannot equal Gamma=7/2 for(2,2), or -1/2 for(2,-2).

In the two-shared-line case, values(2,2) require costs5,8: F6 is
incompatible with equal values and the proved shared-line F7 law
excludes equal values. The F5 has marginal2 and no stars; the F8
would need marginal3/2 and actual star mass1. With two N=-2 at distinct
positions, F8 has either zero glues and two stars, or two glues and
one star. The first has marginal0. For the second, writing
S=l-(3/2)1+(2,2,0) shows that making any two star positions zero forces
all three glue positions nonzero. Thus it is not F8.

Values(2,-2) require costs6,7. The F6 has marginal0 and no stars.
The F7 must have marginal-1/2 and actual star mass1. Its three-glue,
zero-star option fails the mass. Its one-glue, one-star option also
fails: inserting a single glue-1/2 into
S=l+(1/2)1+(2,-2,0) always leaves at least two stars. These finite
scalar assertions are checked exactly for every position choice.

### 1.3 A single class at f11 and the outside-prefix entry B=3

For a sole class a=+2 or-2 at f11, all marginal lies on one F7 main
line, and every B entry toward a vertex outside its prefix is either
0 or1-2a, hence respectively{0,-3} or{0,5}. Neither contains3.

Three main N already use three lines and cost12. Four N could only
be the main pair plus a balanced pair. If their prefixes differ,
they occupy at least three lines. If prefixes agree, two shared lines
have different V positions and cost at least5 each; at total11 their
F5/F6 forms have no actual stars, contradicting mass1. Thus there are
only the two main N. F5+F6 cannot give Gamma=a-1/2, so the costs are
F4+F7, with all marginal Gamma on the F7. A vertex outside the source
prefix belongs to at most one main half, proving the entry assertion.

A related local fact will be used with f<=12. Suppose a sole class
has value b=+/-1/2 and some outside-prefix B entry equals3. Its one
relevant main half would have marginal-3/2. To account explicitly for
balanced N: three main N at this budget leave only glue-free F4 lines.
Four N with different prefixes occupy at least three lines, again
leaving only F4 lines. If the main pair and a balanced pair instead
share two lines, those lines have different V positions and cost at
least5 each. The two shared lines have N entries (-b,t) and (-b,-t), in distinct positions, with b=+/-1/2 and total cost at most twelve. A target marginal -3/2 cannot use F5 or F6. If the target line has F7, its partner must have F5; the partner law forces t=b after orientation. The target then has opposite coefficients, whose F7 marginal is +/-b, not -3/2. If the target has cost at least eight, the pair costs at least thirteen. Thus the shared-pair case is impossible. Five N
plus a star and a nonzero glue already exceed12. Thus the remaining
case has just the two main N. It cannot use an extra line:
two main lines plus a nonzero-marginal third cost at least13. Thus it
contains a specified main N=-b. The line costs at least7 by the
F4/F5/F6 laws. The other main line costs at most5, but would need
marginal Gamma+3/2=b+1, which is neither0,b nor b/2. Contradiction.

If instead b=+/-1 and f<=11, the same argument leaves other-line
cost4 and marginal0. Hence B=3 forces b=-1, f11, and F4+F7 with
F7 marginal-3/2. Here a balanced pair with distinct prefix would need
a third line; with equal prefix the only two shared-line costs within11
are F5/F5 or F5/F6, with no actual stars, contradicting center mass1.
The F7 therefore contains only its main N=+1.
If every crossing glue is in{0,+/-1/2}, marginal-3/2 forces three
glues-1/2, giving two actual stars and cost9, not F7. Thus this last
pattern is impossible under that forced crossing grid.

### 1.4 Two one-loop Z blocks have equal coefficient magnitudes

At C2 with Z in distinct blocks, each block has one L3 and
delta_i=s_i*Z_i/2, s_i in{+1,-1}. Put rho_i=e_partner-e_attachment.
Projection reads

    Z_1*(s_1*c_1-rho_1)+Z_2*(s_2*c_2-rho_2)=0.

Each parenthesized vector has six entries in{+1,-1}, three of each:
c_i occupies the four coordinates outside the source and rho_i its
two source coordinates. Nonzero cancellation of two such vectors
forces |Z_1|=|Z_2|. D does not enter this equation.

### 1.5 Two half-valued Z blocks with no other flow

If Z_1=Z_2=1/2 attach at distinct vertices in distinct blocks and no
other block flows, projection forces either both deltas zero, or
|delta_1|=|delta_2|=1/4. In the zero case the sources coincide and the
attachments are opposite endpoints. This is an exact finite incidence
lemma: all45 blocks and both endpoints are exhausted, and the two
arbitrary rational deltas are solved by their nonsingular2x2 Gram
system and checked in all six coordinates. No delta grid is imposed.

For |delta|=1/4, the net helper-minus-divergence contribution of one
such Z block is+1/2 on its source and one side, and zero elsewhere.
For equal-source zero flow, F on that source has H0=0 at glue-free
endpoints and total helper1, incompatible with w in{0,2}.

## 2. Remove one Z and high helper counts

If z1,d0, Pi>=20. Without cancellation A>=3 and C>=6 gives R>=29.
Cancellation requires Z of coefficient 2, Pi>=21 and A>=2, again R>=29. If z1,d1,
Pi>=20,A>=2,C>=6 would force Pi20. Both endpoints would then be f8,
with the sole class value1/2 at each. The endpoint without Z forces
D=1/2, and the other would force Z=0, a contradiction. For d>=2,
the two endpoints each have two classes and Pi is already too large.
Therefore z>=2 and C>=2.

If any endpoint has two classes, Pi>=21+h. Thus h>=5 exceeds28
with A>=1,C>=2. If no endpoint has two classes, all helpers occupy
one block and h<=3. At h4 the only possible budget is Pi25,A1,C2.
A1 requires K to cross two M-edges and both M-Z values2. The other
two helpers lie outside these blocks and create a new nonzero class
at some endpoint after normalization. One endpoint therefore costs13,
the other at least11, giving Pi>=26. Thus h<=3.

At h3, if both endpoints have one class, all helpers are D+Zu+Zv in
one block. Pi>=21,A>=2 gives C<=5. The one-block lemma makes the Z
equal with zero flow, and the exact kernel deletes both Z while
modifying the existing D and adding at most one E. Support decreases,
contradicting lower bound 31. If an endpoint has multiple classes,
Pi>=24. A1 is impossible: two M cancellations use Z of coefficient 2 at both crossing
endpoints and the third helper creates an outside nonzero array.
Hence only Pi24,A2,C2,X0 remains, with endpoint cost total22.

### 2.1 h3, three Z

Three Z need at most two blocks because C2. A zero-beta outside block
would be an opposite Z pair needing two loops besides the remaining
Z block's loop. Thus all occupied Z blocks have nonzero beta. A2
requires two sole M-Z of coefficient 2 cancellations plus the third Z outside, or
would require an impossible outside zero beta. In the former case
one endpoint has two classes and the other a value2 class, costing
at least13+11>22. This excludes three Z.

### 2.2 h3, one D and two Z

Both Z in one block either share D, already excluded by the kernel,
or give two classes at both endpoints, exceeding the budget. Both Z
in distinct M-blocks are excluded by projection with C2. Both outside
leave too many arrays for A2. Thus one Z is in M and one outside.

If D shares the M-Z block, its other two M-arrays plus the outside
array require A3. A different M-block for D is impossible with two
attachments: its source would be K, and the M-edge at either of its
endpoints is that same block. Therefore D is outside M and must share
the outside Z block with beta0. The M Z is2 and the outside Z is-2D.
Section1.4 gives |D|=1. At the shared M-Z endpoint the class values
are2 and+/-1; section1.1 makes its cost at least14. The other endpoint
has a class of absolute value1 and costs at least9. Their total exceeds22.

Thus only d0,z2 remains.

## 3. Two Z in one block, or at one attachment

For a common block, Pi>=20. If A>=3, C<=5 immediately. A2 would
require beta2 in M and endpoint joint cost at least18, giving Pi>=21
and again C<=5. The one-block lemma makes Z_u=Z_v=a and flow zero.
Source balance and its purely local E23 endpoint argument give joint
cost at least23, hence Pi>=24. An outside block has A>=4, and a
nonunit block in M has A>=3; both exceed28 with C>=2. If a=1 in M,
the local U25 endpoint bound gives Pi>=25, also too large with A2,C2.
The E23/U25 arguments use the common class value and source balance;
they do not require that it was produced by a D instead of equal Z.

Now put both Z at one attachment but in distinct blocks. That center
costs13 and the other, helper-free active center9, so Pi>=23.
At most one Z block is in M, giving A>=3. Only Pi23,A3,C2,X0 remains.
The M Z must be2 to cancel its array; section1.4 forces the other Z
to be+/-2. The active center is f13, excluded by section1.2.

It remains two distinct blocks and two distinct attachments, each
with one nonzero class. Pi>=20.

## 4. Both Z blocks outside M

A>=5 gives Pi<=21,C<=3. If Pi20, both endpoints are f8 with Z=1/2
and all inactive f5. No Z-free block can flow. Section1.5 gives either
the impossible equal-source zero-flow case, or deltas of magnitude1/4.
Each outside group contains at most one M-edge. The two net helper
contributions therefore reach at most two M-edges, each by1/2.
At an untouched M-edge, F would require H0=2, while H0<=1. This
also covers the third-loop and mixed-column slack.

If Pi21, C2 is forced. Section1.4 gives equal absolute class values.
Endpoint total cost is at most18 and each endpoint f<=10. The low-cost
class list leaves either both values1, or values of absolute value1/2
with at most one negative endpoint. An absolute-value1 negative class
already costs10, exceeding18 when paired with the positive class cost9.
Inactive vertices have f<=7 and B entries zero/one.

For values1 the active B entries are nonpositive. A value1/2 center at
f<=10 has all marginals zero. A value-1/2 center costs10 and has only
its two main lines, each F5 of marginal-1/2, so its B entries are at
most2. It is the only possible such endpoint. Consequently every H0
entry is at most3/2, including its edge to the other glue-free active
center. Each outside one-loop Z contributes to at most one M-edge,
leaving an untouched M-edge that would require H0=2. Contradiction.

## 5. Both Z blocks in M

They occupy distinct M-edges U,V; call the unaffected M-edge W.
Projection requires an outside-M flow block, costing two more loops,
so C>=4. Every such outside flow vector must be constant on W.
The proved exact incidence says its matching contains W and neither
U nor V; it therefore has no pair-node divergence at U or V.

With no array cancellation, Pi>=20,A>=3. The possible budgets have
Pi20 or21 and C4 or5. With one cancellation, Z of coefficient 2 gives Pi>=22 and
only Pi22,A2,C4 remains. With two cancellations, only Pi23,A1,C4
remains and both Z are2. These include all mixed/extra-A slack.

### 5.1 C5

Only Pi20 without cancellation fits. Both Z are1/2 at glue-free f8
endpoints, with inactive f5. There is exactly one outside flow block.
At either affected M-source, H0<=1/2 and helper weight3/4, so its
divergence is at most-3/4. Only the other M-Z block can provide that
divergence. A one-loop Z=1/2 block has |delta|=1/4, so each of the
two Z blocks would need at least two loops, in addition to the outside
block's two. This requires C>=6, not5.

### 5.2 C4 without cancellation

Each Z block has one L3. No endpoint can have a separator, since
f<=10, and inactive B<=1. At source U, with Z values a,b,

    w_U <= (1-a)+a+b/2+|b|/2 = 1+max(b,0).

Thus b>=1, and similarly a>=1. The low-cost class list forces a=b=1,
Pi21, active f9 and all inactive f5. F at U,V then forces divergence
-1/2 at each and B from each inactive mate to its active mate equal1.
The two M-flows contribute+1 at W. Projection on W forces the outside
flow to contribute-1 there and+1 at another pair-node e. At e, F
requires H0_e=1. An edge incident to an active f9 center has H0<=1/2,
so e must join the two inactive mates p,q of the Z attachments u,v.

But the private f9 lines force p,q into opposite halves. To see this,
the u line contains U={u,p} and a third vertex t. It is dense. If
t=q, its glue to v lies on a forbidden separating v line. If t=v,
a positive u-glue forces an inactive opposite f5 line to be dense,
which activates the likewise forbidden separating v line. Thus t
belongs to W. The opposite half then contains V={v,q} and the other
W endpoint. Its v line is dense and forces all lines in that partition
to occur. Hence p,q lie in opposite halves and H0_pq=0, not1.

### 5.3 One or two cancellations

With one cancellation a=2, Pi22 leaves the other endpoint f<=9.
The preceding source inequality still forces b>=1, so b=1. The exact
endpoint costs are11,9 and all inactive f5. At the unit source V,
H0_V is-1/2 or0 and helper weight is2, so divergence is-1/2 or0.
But its only contributing flow is the one-loop Z of coefficient 2 block, whose
divergence there has magnitude1. Contradiction.

With two cancellations a=b=2, both endpoints are f11 and all inactive
f5. At either source H0 is-3/2 or-1 and helper weight3, so divergence
is again-1/2 or0, whereas the other one-loop block gives magnitude1.
This completes both-in-M placements.

## 6. One Z block in M and one outside

Let M-Z=a attach at u with source e={u,p}; let outside Z=b attach at v.
First suppose C<=3, so no Z-free block flows. If neither endpoint has
a separator, classify e in the outside group:

- If e is its source, p=v and F gives w_e=(1-a-b)+(a+b)=1, impossible.
- If e is a side, both flow vectors are constant on e and the outside
  source is disjoint from it, contradicting the antisymmetric M-Z term
  in projection.
- If e is absent, F requires B_pu=3, since B_up=1-2a and e has no
  outside helper contribution or divergence.

The following budgets ensure the stated no-separator premise and
exclude the final B_pu=3 possibility.

### 6.1 A3

The M array must cancel, so a=2, Pi>=22, and Pi<=23. At Pi22,
f(u)<=12, f(v)<=9, and every inactive center is light. If p is inactive
its B entry is at most1; if p=v, the proved outside-prefix f<=10
bound gives B<=2. At Pi23 the budget forces C2. Section1.4 gives
|b|=2, making both active costs11 and all inactive f5. There are still
no separators. If p=v, section1.3 gives B_vu in{0,-3} or{0,5}, not3.
This includes Pi22,C3 and all mixed slack.

### 6.2 A>=4 and C<=3

Pi<=22. If Pi<=21, both active f<=10 and inactive f<=7, so the same
outside-prefix bound excludes3. At Pi22, only A4,C2 fits. Both active
f<=12, all inactive are light, and section1.4 gives |a|=|b|. The only
remaining possibility is p=v with f(v)>10. Then f(u)<=9:

- f(u)=8 forces a=1/2, b=+/-1/2, f(v)<=12, excluded by section1.3.
- f(u)=9 forces a=1, b=+/-1, f(v)<=11. Section1.3 leaves only b=-1,
  f(v)=11, an F7 main line of marginal-3/2. All inactive centers are
  f5. That line's half contains both active vertices u,v and its
  outside source partner, so all opposite centers are inactive f5.
  Their forced half-grid makes this F7 impossible by section1.3.

### 6.3 C4

Only Pi20,A4,C4,X0 fits. Both Z are1/2 at glue-free f8 endpoints and
all inactive f5; hence H0 is nonnegative and at most1 everywhere.
If no Z-free block flows, section1.5 gives the already impossible
equal-source zero-flow case, or each Z contributes net1/2 at its source
and one side. If a Z-free block flows, it uses two loops, leaving one
L3 at each Z block and giving the same net contributions.

The M Z contributes to two M pair-nodes; the outside Z contributes
to at most one. Their total contribution on M is at most3/2, and each
M node receives at most1. Thus H0+net_helper-2 is nonpositive on M
and strictly negative at at least two M nodes. It is nonnegative off M.
With no further flow this is impossible. With the one Z-free flow,
it is also impossible: a single graph edge has only one negative
divergence node. This treats all idle-loop placements in the C4 budget.

These cases exhaust two Z in different blocks and at different
attachments. Together with sections2--3, they close every m2 Z-present
matching R28 case.
