Assume matching corrections, R=28, Z present and m in {3,4}.

Use the lower bound 31 identities, normalization and local costs.
Set h=d+z, R=Pi+A+C+X, 2Pi=sum f+2h. Inactive f>=5 with no f6;
active f>=8, or9 if Gamma!=0. Two nonzero classes give f>=13. An
active f8 has exactly one nonzero class of value1/2. At f9 a nonzero
class has value1; otherwise all helpers vanish. A sole value2 class
costs at least11; if another class is present f>=13, so the bound11
remains valid. The helper-array charge is Pi>=21+h for m4 and
Pi>=20+h for m3. Matching rigidity gives Omega=-2e_M.

Every Z block needs a loop. Exactly one global Z needs C>=6. With Z
present, C1 is excluded by the proved support-decreasing normalization
against lower31, so C>=2. A Z-free nonzero-flow block costs two loops.
Mixed columns vanish in all equations used here.

## 1. Cancellation costs and projection facts

If an M-array cancels, its block has beta=zeta_u+zeta_v=2. If only one
source endpoint is active, it has class value2 and costs f>=11. If
both are active, their joint cost is at least18. Indeed f<=9 permits
the relevant class values only0,1/2,1, with values1 costing9, and
f8+f10 cannot sum to beta2: the full f<=10 single-class list is
{1/2,1,-1/2,-1}, while a second class already costs13. At joint cost18,
the only cancellation pattern is costs9,9 and both relevant values1.
All these statements allow balanced N and additional classes.

For flow blocks with group M, the vector c=1_W-1_V is constant on
each M-edge. A lone endpoint Z on an M-edge gives a nonzero
antisymmetric component there in the projection equation. Consequently
all flow confined to M cannot coexist with any unpaired endpoint Z
in an M-block. If two distinct Z blocks use C2, every loop is in them.
If C<=3, no additional Z-free nonzero-flow block can occur.

The proved one-block lemma says that if all Z form the endpoint pair
of one block and C<=5, the two Z are equal and all flow is zero. If
a D already occupies that same block, the exact move

    (D,Zu,Zv,E) -> (D+t,Zu-t,Zv-t,E+t), t=Zu=Zv,

removes both Z, adds at most one E and changes an already present D.
It strictly decreases support while preserving all classes and the
correction terms, contradicting the proved global lower31 at width31.

## 2. Four attachments

Pi>=23 excludes z1 immediately, since C>=6. Hence z>=2,C>=2.

### 2.1 A0

If K consists of two full M-edges, the third array cannot cancel.
Otherwise K contains one full edge U and one endpoint of each other
edge. Canceling the two crossing arrays needs two class-value2
vertices, costing at least11 each. The other two active vertices cost
at least8 each and the inactive vertices at least10 together. At least
three helpers are needed to cancel all three arrays, so

    Pi>=ceil((11+11+8+8+10)/2)+h=24+h>=27.

With C>=2 this exceeds28. Thus A>=1 and Pi<=25, whence h<=4.

### 2.2 A1

At least two M-arrays cancel. If h4, Pi25 is forced, and every active
vertex is f8 and every inactive f5. All relevant class values are1/2,
so no block has beta2. If h2, the helpers are two Z of coefficient 2 needed in two
canceled blocks. At distinct endpoints the active cost is at least
11+11+9+9=40; at one common endpoint it is at least13+9+9+9=40.
Thus Pi>=27, impossible.

For h3 there are z=3 or d=1,z=2. With z=3, at least one canceled block
has a sole Z of coefficient 2, costing11 at its endpoint, and at least one active
vertex is helper-free and costs9. The remaining two cost at least8,
giving Pi>=ceil((11+9+8+8+10)/2)+3=26.

For d=1,z=2, distribute the three helpers among the two canceled blocks.
If one canceled block uses D plus Z, the other has a sole Z of coefficient 2. At most
three active vertices carry nonzero classes, so the same11,9,8,8 bound
gives Pi>=26. If one canceled block uses a Z pair and the other a
sole unit D, disjoint source endpoints cost at least18+18=36; intersecting
sources give a two-class vertex and a helper-free vertex, which only
increases that bound. Again Pi>=26.

It remains two cancellations each using one helper, with the third
elsewhere. Sole unit D and sole Z of coefficient 2 require active cost at least37 if their
endpoints are distinct, or at least13+9+8+8=38 if they overlap. Two
sole Z of coefficient 2 require cost at least38 at distinct endpoints; if they share
an endpoint, that vertex costs13 and a helper-free vertex costs9,
with the other two at least8 each, again38. Thus no h3 case fits Pi25.
These are all allocations; an outside zero-beta block needs two helpers
in addition to the two cancellations and is covered by the h4 exclusion.

Therefore A>=2, Pi<=24 and h<=3.

### 2.3 The last h2/h3 budgets

At h2 the helpers are two Z. If both attach at one center, the active
cost is at least13+9+9+9, already too large. Otherwise the other two
active centers cost9 each, so Pi>=ceil((8+8+9+9+10)/2)+2=24.
The only budget is Pi24,A2,C2,X0. Its two Z endpoints are f8 with
values1/2; no M-array cancels, contrary to A2.

At h3,z=3, at least one active center is helper-free, giving
Pi>=ceil((8+8+8+9+10)/2)+3=25. At h3,d=1,z=2, Pi24 forces all four
active f8 and inactive f5. Every relevant class is1/2, so beta<=1
and again no array cancels. This completes m4 with Z present.

## 3. Three attachments: remove small A and large h

A single Z has Pi>=22: with d0 the other two active vertices cost9
each; with d>=1 use Pi>=20+h. At least one M-array remains untouched,
so A>=1. Together with C>=6 this gives R>=29. Thus z>=2,C>=2.

### 3.1 A0

K must meet every M-edge once. Each canceled array then needs Z of coefficient 2 at
its active endpoint. Active cost is at least33, inactive cost15 and
h>=3, so Pi>=24+h>=27. Hence A>=1, Pi<=25 and h<=5.

### 3.2 A1

If K meets all three M-edges once, two distinct vertices have Z of coefficient 2.
At h>=3, costs11,11,8 and inactive15 give Pi>=26. At h2 the third
active vertex is helper-free; only Pi25,A1,C2,X0 remains. Both Z
are lone endpoints in distinct M-blocks, contradicting projection.

Otherwise K=U union{w}, where U is an M-edge and w lies on another
M-edge V. The fully inactive third edge consumes the one alternating.
Both U and V cancel; V has its sole Z_w=2 and f(w)>=11. At h2,
z>=2 forces a sole Z of coefficient 2 at an endpoint of U, with the remaining active
vertex helper-free. This again gives Pi25,A1,C2 and the projection
contradiction.

At h3, all helpers lie in the two canceled M-blocks: an outside
nonzero beta would require another alternating, and an outside zero
beta would need two extra helpers. V contains only Z_w. Cancellation
at U costs joint f>=18, so Pi>=ceil((18+11+15)/2)+3=25. Equality
requires both U-class values1 with endpoint costs9,9. D+Z would make
the two endpoint values unequal unless the Z were zero. Thus U has
Z=1,1, and V its lone Z of coefficient 2. C2 confines loops to M, again impossible.

At h4, Pi>=ceil((8+8+11+15)/2)+4=25. Equality forces the two U
endpoints f8, so their relevant values sum to at most1, not beta2.
h>=5 is already too costly. Consequently A>=2, Pi<=24 and h<=4.

If h4, the line budget is at most40, with active total at most25.
A2 requires a cancellation. A cancellation within K costs18 plus
the third active cost8; a crossing cancellation costs11+8+8. Both
exceed25. Thus h4 is impossible, and only h2 or h3 remain.

## 4. Three attachments, h2: two Z and no D

Two Z at one attachment cost at least13 there,9 at each other active
center and15 inactive, giving Pi>=25. Therefore their attachments
are distinct; the third active vertex is helper-free. Pi>=22.

### 4.1 Both Z in one block

A>=2 implies C<=4. The one-block lemma gives equal Z and zero flow.
Source balance requires a nonzero separator at one endpoint. Endpoint
costs at least13,8, third active cost9 and inactive15 give
Pi>=ceil((13+8+9+15)/2)+2=25, impossible.

### 4.2 Distinct blocks, both in M or both outside M

Both outside requires A>=5 and R>=22+5+2=29. Both in M requires an
additional Z-free flow block by projection, so C>=4. Without array
cancellation, A>=3 already gives29. One canceled array has a Z of coefficient 2 and
forces Pi>=ceil((11+8+9+15)/2)+2=24; with A>=2 this gives30. Two
canceled arrays force Pi>=25 and A>=1, again giving30.

### 4.3 One M Z=a at u, one outside Z=b at v

If A=3, cancellation forces the matching Z to have a=2. The active costs are at least 11,8,9 and the inactive cost is 15, so Pi>=24 and R>=24+3+2=29, outside the present bound. Thus A>=4. The only remaining budget is Pi22,A4,C2,X0; both Z endpoints have f8 and a=b=1/2, the helper-free active center has f9, and the inactive centers have f5.

Write e={u,p} for the M source. The M endpoint has no nonzero
separator, so B_up=1-2a. All flow is confined to the two Z blocks.
Classify e relative to the outside matching group:

- If e is its source, p=v. Neither endpoint has a separator, so
  H0_e=1-a-b. Helpers contribute a+b and divergence is zero, giving
  w_e=1 rather than2.
- If e is a side, both flow vectors are constant on e and the outside
  source is disjoint from e. The M Z gives a nonzero antisymmetric
  projection on e, a contradiction.
- If e is absent, it has no divergence or outside helper contribution.
  The F equation forces B_pu=3. An inactive p has B<=1. At the outside
  Z center, u is outside its prefix and the proved f<=10 bound gives
  B<=2. At the helper-free center, the new B10 forms give B<=2:
  at f9 entries are0/1; at f10 they belong to{0,2,-1,1} because the
  two marginals are-1 and1/2. None can equal3.

This includes the extra helper-free attachment without assuming its
B entries are always nonnegative. It closes h2.

## 5. Three attachments, h3

Pi>=23. At Pi23 the active line total is25, so there are two f8
vertices and one f9, with all inactive f5. No vertex has two nonzero
classes. Every nonzero class is1/2 or1 and no block can have beta2;
thus A>=3 and the budget is Pi23,A3,C2,X0. Every nonzero helper
block has positive beta, so all helpers must lie in M.

With three Z, they occupy at least two blocks and some block has a
lone endpoint Z. C2 confines flow to M and projection excludes it.
With one D and two Z, the D is on the sole full M-edge in K. The Z
either both lie in its block, excluded by the equal-Z kernel reduction
of section1, or one lies at the third active vertex in another M-block.
The latter is a lone Z, excluded by projection with C2. Thus Pi23 fails.

At Pi24 the exact budget is A2,C2,X0 and at least one M-array cancels.

### 5.1 Three Z

There are at most two Z blocks. A zero-beta outside block would contain
an opposite Z pair and need two loops, in addition to the remaining
Z block's loop. Thus every Z block has nonzero beta. An outside Z
block plus at most one canceled M-array would require A>=3. The remaining possibility of two canceled matching arrays and an outside Z must also be excluded. Two value-two matching Z at different attachments, together with an outside Z, cost at least 11+11+8+15 in lines, hence Pi>=26 after three helpers. If the outside Z shares an attachment, the two-class center costs at least thirteen and the line charge is at least 13+11+9+15, hence Pi>=27. Both exceed Pi24. All Z therefore lie in M, with a lone endpoint Z in some block, contradicting projection with C2.

### 5.2 One D and two Z in one block

The one-block lemma gives equal Z and zero flow. If D is in that
same block, the kernel reduction contradicts lower31. If D is in a
different block, both blocks have nonzero beta. A2 forces both into M;
their disjoint source pairs would require four attachments, not three.
This case is impossible.

### 5.3 One D and two Z in distinct blocks

Both Z in M are excluded by projection with C2. Both outside cannot
fit A2: a single D can cancel at most one M-array or one outside beta,
leaving too many arrays. Thus exactly one Z is in M and one outside.

If D shares the M-Z block, the other two M-arrays and the outside Z
array give A>=3. If D is in a different M-block, A2 requires D=1
and M-Z=2. Their endpoints exhaust K. The outside Z adds a second
class somewhere. At a D endpoint the active costs are at least13,9,11;
at the value2 endpoint they are at least9,9,13. Either exceeds the
Pi24 active budget27.

If D is outside M, the M Z must be2, and the outside beta must vanish.
Thus D and the outside Z share a block, Z=-2D, and its endpoint
classes are nonzero opposites. If its source contains the M-Z endpoint,
that center has two classes and f>=13, the other source endpoint
costs at least8, and the third active vertex is helper-free with f>=9.
This gives Pi>=26. Otherwise the M-Z endpoint costs11 and the opposite
classes cost at least8+9, giving Pi>=25. Neither fits Pi24.

This exhausts m3 with Z present. All uses of C0 or confined flow have
been justified by a loop budget or the one-block lemma; zero flow alone
has never been used to discard idle loops.
