## 1. Equations, stated local facts, and exact scope

Write M={01,23,45}, K the three vertices with N support, h=d+z, and

    R=Pi+A+C+X,       2(Pi-h)=sum f,
    f=g+2s+2N,        Gamma(x)=sum_B zeta_x(B)-1/2.

s counts actual stars, whose coefficient sum at each center is1.
For a block B=(G,U={u,v}), endpoint classes are zeta_u=D+Z_u,
zeta_v=D+Z_v, and beta=zeta_u+zeta_v. Alternating array totals are

    t_U^G=-2*1[G=M]+beta_B.

The mixed-free necessary equations are

    H0_uv=(B_uv+B_vu)/2,
    B_uv=-2 sum_(u-halves containing v) r,
    div(delta)_e=H0_e+sum_B beta_B*(1_U+1_V/2+1_W/2)-2*1[e in M],
    2 sum_B delta_B c_B=sum_Z Z*(e_mate-e_attachment),
    c_B=1_W-1_V,      S=l-r*1-nu.

Each nonempty zero-beta alternating array needs at least two columns;
each nonzero array needs at least one. Every source of a surviving D
lies entirely in K; its two endpoint classes are nonzero after the
normalization. Every nonzero Z has an attachment in K.

The lexicographic (R,z,d) representative has no equal nonzero endpoint
Z pair and satisfies C≥z. A Z-free block with nonzero flow costs at
least two loops. A block with one Z has |delta|=|Z|/2 when it has just
one loop. Idle loops are included in C and cannot create an additional
flowing block for free. The earlier one-block lemma removes a
sole global endpoint Z pair at m3: A≥2,Pi≥22 gives C≤5, where equality
of the two Z would contradict the chosen normal form. We use that result
only in this stated scope. The z1 branch was already excluded.

Inactive centers cost at least5, have no f6, and their strict excess
starts at2. At f≤9 their sole nonzero marginal is-1/2 and B is0/1.
Active centers cost at least8; a helper-free active center costs at least9.
At f8 an active center is glue-free and has a single nonzero class1/2.
At f9 a nonzero class is1; all glues have magnitude1/2. Helper-free
active f9 has B in{0,1}.

Two nonzero classes cost at least13. Three cost at least14. The proved
T13 fact excludes a value2 class at f≤13 if another nonzero class is
present. In its present uses the added value2 prefix differs from the
incident D prefix, or the value2 obstruction is applied as previously
proved. No isolated shared-F7 shortcut is invoked.

For a sole nonzero class at f≤10 the signed value list is
{1/2,1,-1/2,-1}; exact f10 has no value1. A center of such a class has
B≤2 when queried outside its source prefix. Nonzero marginal separation
from that prefix costs at least13, so if f≤12,

    B_(source mate)=1-2*(class value).

Helper-free active f≤10 has B in[-1,2]. This is the balanced
N law, not inactive lightness. A sole value2 class costs at least11.
Cancellation of an M-array with two active source endpoints costs joint
f at least18; at18 both endpoint values are1, with f9/f9. With one
active source endpoint, cancellation needs its lone Z of coefficient 2 and f≥11.
All these local facts include balanced N and arbitrary signs.

## 2. Two small new lemmas

### 2.1 Scoped balanced B11 and half-grid parity

Suppose a helper-free active center w has f≤11, the other two active
centers are glue-free, and all three inactive centers are F5. Every
glue at w then crosses an inactive F5 line and equals ±1/2. Its Gamma
is-1/2. **Every directed B entry at w is at most2.**

Only the f11 boundary is new. Balanced helper classes have an even N
count: each h-side of each class has exactly two possible N columns;
zero class sum makes both present or both absent. Their two columns
occupy different private lines. Since Gamma and actual center mass require
a glue and a star, N is2 or4. At N4, g1 and s1 are forced, so B≤1.
At N2, g is1,3 or5. The g1 and g3 possibilities have at most one and
two negative half glues, respectively, hence B≤1 and B≤2.

The g5 case has s1, with the sole actual star equal to1. All glues are
on the two N lines; an extra line would cost at least5+4+4=13. Five
glues divide as3+2. A star-free two-glue half-grid line cannot have just
one nonzero aggregate-N cell. Thus the three-glue line is star-free;
its F5 one-N law forces its N amplitude to have magnitude1. On the
two-glue line, r is an integer and the two nonzero glue positions have
half-integer effective stars. Integer N and the actual star1 cannot
supply those entries. Contradiction.

A direct exact control checks all3240 masks: the missing glue, two positive
glue positions, both N positions, and the location of the actual star.
The N amplitude is reconstructed from one cell equation as an unrestricted
rational root, and every other cell is checked. There are no survivors.

We also use the simpler parity consequence of half-grid glues:

    f ≡ g ≡ 2Gamma (mod2).

Thus a sole class3/2 at f11, a sole class-1/2 at f11, or a sole class-1
at f10 is impossible when all its crossing glues are forced onto that
half grid. Each application below first proves that all other possible
crossing centers are either glue-free or have half-magnitude glues.

### 2.2 A matching-edge aggregate for the half-Z equality cores

If two Z endpoints are f8, the third active center is helper-free f9,
and all three inactive centers are f5, then

    sum_(e in M) H0_e ≤2.

The two f8 centers have B rows zero; the other four centers contribute
at most1 each to their own matching mate. Summing H0 counts each of
these six directed entries with factor1/2.

Distinct perfect matchings share at most one edge. For one Z=1/2 in an
outside-M block with one loop, |delta|=1/4. Its total helper contribution
minus flow divergence on M is at most1/2: if its source is in M the
helper term is1/2 and divergence there is zero; if one side is in M
the helper term is1/4 and flow correction at most1/4; otherwise it is0.
Two such outside blocks therefore give

    6=sum_M H0+helper-divergence ≤2+1=3,

an exact contradiction.

If instead both half-Z blocks are in M, each contributes total1 on M,
with matching-summed divergence zero. Nonzero antisymmetric projection
forces an outside Z-free flowing block. At C4 this is the only extra
flow block: the two Z blocks each consume one loop, and it consumes two.
On each half-Z source M-edge the right-side endpoint difference has
absolute value1, while all M flow vectors are constant there. Hence

    2|delta_out|*|c(u)-c(mate)|=1,
    |delta_out| in {1/4,1/2}.

An outside block contributes divergence on at most one M-edge, so

    6≤2+1+1+1/2=9/2<6,

again impossible. The fresh exact projection check covers all504 source,
endpoint and outside-block configurations.468 are inconsistent; the36
compatible six-coordinate systems force the stated bounded flow.
The aggregate works even if extra idle loops are admitted before the
cost equality is established. It does not assume star positivity at
unrestricted active centers.

## 3. A0, repeated D and large helper counts

### A0

All three M-arrays must cancel. K must meet each M-edge exactly once,
because otherwise one M-array has no possible active helper endpoint.
Every cancellation is therefore a lone M-Z of coefficient 2. The active costs are at
least33, inactive costs15, and h≥z≥3. Thus

    Pi≥24+h,       R≥24+h+z≥30.

So A≥1.

### d≥2

For d≥3, the source multigraph on three vertices has at least two
vertices of D degree at least2. Each has at least two distinct nonzero
classes, so the active sum is at least13+13+8=34. Inactive sum15 gives
Pi≥25+h≥30 because h=d+z≥5. This uses no Z-free degree lemma.

For d2 with repeated source, the same34 bound gives Pi≥25+h≥29,
already impossible with C≥z≥2.

For d2 with different sources, the D graph is a path. Its shared center
costs at least13 and the leaves at least8, giving Pi≥22+h. If z≥3,
R≥24+2z≥30. If z2, Pi≥26,C≥2 forces A≤1, hence A1. Two canceled
M-arrays include one with only one active endpoint. Its value2 class
is additional to that endpoint's incident D classes, since the M-mate
is outside K. At the shared center three classes give f≥14 and
Pi≥23+h=27; at a leaf T13 gives f≥14 while the shared center remains
at least13, a stronger bound. Either yields R≥30. Thus **d≤1**.

### A1 and z≥4

Two M-arrays cancel. If K meets all three M-edges once, the active sum
is at least11+11+8=30, giving Pi≥23+h. If K contains one full M-edge
U and one endpoint of another V, the U cancellation and lone V-Z of coefficient 2 give
active sum at least18+11=29, giving Pi≥22+h.

With d≤1 and z≥4, or d1,z3, these bounds and C≥z already exceed29.
For h2 the helpers are two Z of coefficient 2 in distinct M-blocks and Pi≥25, hence
C≤3. For h3 the two cancellations consume at least two helpers. An
outside zero-beta block needs at least two more helpers, and an outside
nonzero-beta block would need an additional alternating. Thus all helpers
are in M and Pi≥25, again C≤3. A Z-free outside flow block would need
two loops in addition to the Z-block loops, so none fits.

There is an unpaired M-Z: at z2 the sole-pair-in-one-block
exclusion leaves distinct blocks; at z3 the odd number leaves a lone
endpoint. Its antisymmetric right-side projection cannot be supplied
by flow vectors constant on every M-edge. This excludes A1.

Now A≥2 and

    R≥(20+h)+2+z=22+d+2z.

All z≥4 disappear. Only d0,z2; d0,z3; d1,z2; and d1,z3 remain.
At d1,z3 the sole possible tuple is Pi24,A2,C3,X0. Its active sum is
at most25, so no center is multiclass, and its class values are1/2 or1.
A crossing cancellation needs a value2, while an internal cancellation
requires two f9 centers and the third costs8, exceeding25. Thus A≥3,
a contradiction. We have reduced to the three families below.

## 4. Three Z, no D

A≥2,C≥3 force Pi≤24 and active sum≤27. A multiclass center would
cost13 plus8+8=29, so the three Z attach at distinct vertices, one
class each.

At Pi23 the active sum is at most25: every class is1/2 or1, every
Z-block beta is positive, and no M-array cancels. Hence A≥3, and
only Pi23,A3,C3,X0 can fit. There is no outside nonzero beta, so all
helpers are in M. Three loops are already required by the three Z;
all flow is confined to M, and the unavoidable unpaired Z contradicts
projection.

At Pi24 only A2,C3,X0 fits. Two M cancellations cost at least29
active units (18+11 for one internal plus one crossing cancellation,
or11+11+8 for two crossing ones), exceeding27. Exactly one cancels;
the two other M-arrays consume both alternatings. Any outside helper
block must have beta0. With D=0 this would use an opposite endpoint Z
pair, whose joint cost is at least18: one endpoint costs at least10,
since f8/f9 nonzero classes are positive, and the other at least8.
The remaining Z must be the sole canceled M-Z of coefficient 2 and cost11. Their active
sum29 exceeds27. Therefore all helpers are in M. C3 leaves no outside
flow, and an unpaired M-Z gives the same projection contradiction.

Thus all d0,z3 cost records are excluded.

## 5. One D and two Z

After A≥2,C≥2, only Pi23,24 or25 is possible. The Z occupy distinct
blocks by the one-block result.

### Pi23: the new A4 boundary

The active sum is at most25, so every active center has one nonzero
class. Exactly one Z shares a D endpoint class; the other Z attaches
at the third center. Otherwise there is a multiclass vertex, or both
Z occupy the D block, already excluded. The D endpoint values are
unequal, because their difference is the nonzero shared Z. They must
therefore be1/2 and1 at costs8,9; the third Z is1/2 at cost8. All
inactive centers are f5. The two helper-block betas are3/2 and1/2,
and no M-array cancels.

At A3 all helpers are in M; C≤3 confines flow to M and gives the
unpaired-Z projection contradiction. At A4 only C2 fits, and at most
one helper block is outside M.

If the D/Z block is in M and the third Z outside, let e be the D source.
It cannot be the outside source, since the third attachment is not a D
endpoint. If e is an outside-group side, both flow vectors are constant
on e while the shared M-Z gives nonzero antisymmetric projection; the
other source is disjoint from e. If e is absent from that group, its
equation requires H0_e=2-beta but the two nonseparating endpoint rows
give H0_e=1-beta. Both are contradictions.

If the D/Z block is outside and the third Z in M, let e={w,p} be the
third Z's source. Again it cannot be the D source. If e is a side,
projection is impossible; if absent, its source equation forces B_pw=3.
But p is inactive f5 or a D endpoint f8/f9, and w is outside that endpoint's
class prefix, so B_pw≤2. This excludes the Pi23,A4 boundary too.

### Pi24: cancellation, and the signed zero-beta boundary

The active sum is at most27, so every center still has one class. The
same D/Z sharing pattern applies. A2 needs an M cancellation. A sole
third M-Z of coefficient 2 would cost11, leaving joint D endpoint cost at most16;
both would be f8 of value1/2, making the shared Z zero. Cancellation
at the D/Z block leaves joint endpoint cost at most19. The low-cost
possibilities are f9/f9 with values1,1 (again shared Z=0), or f8/f11
with values1/2,3/2. The latter forces the other active vertex glue-free
and all inactive vertices f5, so half-grid parity excludes class3/2 at
f11. There is no M cancellation. Hence A≥3.

At A3, C2 is forced. All-M helper blocks fail projection. Otherwise
the third Z must be in M and the D/Z outside block must have beta0;
any nonzero outside beta would add a fourth alternating. Its endpoint
classes are opposite. An opposite pair costs at least18, so all inactive
centers are f5 and the active sum is exactly27. The only signed low-cost
patterns, up to endpoint reversal, are

    (D endpoint f, D endpoint f, third Z f)=(8,10,9),(8,11,8),(9,10,8).

The latter two have opposite classes1/2,-1/2 or1,-1. The other endpoint
has half-magnitude glues or none; the third is glue-free; every inactive
line is F5. Thus half-grid parity excludes the negative f11 or f10
endpoint respectively. The first pattern has values1/2,-1/2,1.
For the third M-Z source e, the same source/side/absent argument applies:
source equality is impossible because the third vertex is not a D
endpoint; a side gives antisymmetric projection; absence forces B=3
at an inactive center or an outside-prefix f≤10 D endpoint, all bounded
by2. This closes Pi24.

### Pi25: allow the multiclass equality, then remove it

Only A2,C2,X0 fits. The active sum is at most29. If a multiclass vertex
occurs, its cost is13 and the others are f8, with all inactive f5.
No Z can share the D block in this case: an extra class at a D endpoint
would leave a helper-free third attachment and active cost at least
13+8+9=30. Therefore the lone D has coefficient1/2 from an f8 endpoint,
and beta1. Any M-Z of coefficient 2 at an f8 vertex is impossible; at the f13 multiclass
vertex it is excluded by T13. There is no M cancellation, contrary to A2.

Otherwise there is one class at each vertex, with one Z sharing the D
block and the other at the third vertex. If the D block cancels an M
array, A2 forces the third Z into M too; all flow is in M at C2 and
projection fails. If the third M-Z of coefficient 2 cancels, an inside-M D block has
the same projection obstruction. An outside D/Z block must have beta0,
since the other two M-arrays consume A2. The third class costs11,
leaving joint opposite-class endpoint cost at most18, hence f8/f10
with values1/2,-1/2. The third center has no separator. Its source e
cannot be the D source; a side violates projection, while absence forces
B=3 at an inactive or outside-prefix f≤10 endpoint. Again B≤2.

Thus all d1,z2 cost records are excluded.

## 6. Two Z, no D

If both Z attach at one center, active costs13,9,9 plus inactive15
give Pi≥25. At most one of their distinct blocks lies in M; even with
cancellation the outside block gives A≥3, so R≥30. Their attachments
are therefore distinct; the third active center is helper-free. A sole
Z pair in one block was already excluded. Pi≥22.

### Both blocks outside M

A≥5, so only Pi22,A5,C2,X0 fits. The two Z centers are glue-free f8
with Z=1/2; the third active is f9 and inactive centers f5. Lemma2.2
gives 6≤3 for the matching-edge sum, impossible.

### Both blocks in M

Unpaired M-Z projection needs an extra outside flow block, so C≥4.
One cancellation gives Pi≥24,A≥2 and R≥30; two give Pi≥25,A≥1
and R≥30. Without cancellation only Pi22,A3,C4,X0 fits. Its half-Z
private equality satisfies lemma2.2, giving6≤9/2, impossible.

### One M-Z=a at u and one outside Z=b at v

With no M cancellation A≥4, so Pi≤23 and C≤3. The M center has
f≤10; the outside center f≤10; inactive centers f≤7; the helper-free
third center f≤11. There is no nonzero separator at either Z center,
and every flow remains in the two Z blocks. Let e={u,p} be the M source.

- If e is the outside source, p=v. Both endpoint rows are nonseparating;
  H0_e=1-a-b. Helpers contribute a+b, divergence is zero, and the edge
  weight would be1 rather than2.
- If e is an outside-group side, both flow vectors are constant on e,
  the outside source is disjoint, and the M-Z antisymmetric projection
  cannot be supplied.
- If e is absent, its source equation forces B_pu=3. An inactive p
  has B≤1; the outside endpoint has the outside-prefix B≤2 law; a
  helper-free p at f≤10 has signed B≤2. At f11 the line budget forces
  both Z centers f8 and every inactive f5, exactly lemma2.1's scope,
  so the new B11 bound also gives B≤2.

With cancellation a=2, A≥3 and Pi≥24, so only Pi24,A3,C2,X0 fits.
The M endpoint has f≤12, the outside endpoint f≤9, the helper-free
center f≤10, and every inactive is f5. The same source/side/absent
argument applies, now using only the f≤10 bounds. No SZ6
or extended sole-Z lemma is invoked.

This exhausts d0,z2 and hence every identity with m=3 and a nonzero Z at R≤29.
