## A. Complete outer cover and noncircular minimality

Assume a PM2 representation with at most30 directions. The proved global
lower30 theorem makes its width exactly30. The proved canonical
completeness theorem and normalizations do not increase support; their
outputs remain PM2 identities, so their width is again30. A subsequent
strictly support-decreasing exact kernel move is impossible by the
preceding lower bound 30. This reasoning never assumes lower31.

Write W=e+R. The canonical correction graph has nonzero signed edge
weights w of coordinate degree2 after the standard scale3 conversion.
No vertex is isolated. Fewer than3 edges cannot cover6 vertices;3 edges
give a perfect matching, with each weight2. Four edges are impossible.
For a graph on six nonisolated vertices with four edges, its component
count c satisfies c>=6-4=2 and c<=3. If c3, all components have size2
and the simple graph has only three edges. Thus c2, and four edges
attain the minimum6-c, so the graph is a forest, of sizes2+4 or3+3. The degree
equations force equal cardinalities of the two bipartition classes in
each component, ruling out odd components and the4-star; the remaining
4-path would have a zero middle edge. This is valid for signed weights.

The proved retained bounds are R>=27 for matching e3, R>=25 for e5,
and R>=24 for every other nonmatching graph. Thus e>=7 already costs
at least31. Exactly the following budgets need exclusion:

| Correction support | Retained support at width30 | Component |
|---|---|---|
| e5 | R25 | C: nonmatching |
| e6 | R24 | C: nonmatching |
| e3, matching | R27 | D-I: matching |

For matching R27, split first by whether all normalized zetas vanish.
Part F covers that case for every attachment count. A nonzero normalized
zeta needs nonzero N on both h-sides at its vertex. Every surviving D
or Z therefore attaches in K, and a nonzero helper forces m>=1.
Normalization changes only D,Z,E; it preserves every zeta, every N,
K,e and the correction weights w, and does not increase R. For the nonzero-helper case, part G
covers m1, H covers m2, I covers m3, and D covers m4,m5,m6.
These are all possibilities; balanced zero-zeta N at any counted
attachment is retained. 

## B. Proved inputs, notation and verification boundary

The complete canonical model covers all22,155 primitive projective rays
of the PM2 dictionary, reducing without support increase to1530 retained directions
and15 correction directions. Its certificates and the already accepted
global lower30 proof are the foundation, not new claims in this document.
The preceding component supplies the lower bound 30.

Use R=Pi+A+C+X, Pi=stars+glues+N+Z+D, C=L3+E, K the N-attachment
vertices, m=|K| and h=d+z. Counts mean nonzero columns, never coefficient
sums. A private line has cost F=g+2s+2n, and f(x) sums its line costs:
2Pi=sum_x f(x)+2h. Inactive f>=5 with no f6; active f>=8, or9 when
Gamma!=0. Actual star mass is1 at every center. These statements permit
signed stars, balanced N and collisions. Private/glue marginal costs
refer to actual selected columns, without charging replaceable stars twice.

The D-degree bound used in D and I is the Z=0 bound in section1 of
the lower-bound-30 component above. For incident D-degree
d_x>0, pi(x)=f(x)/2+d_x/2 is at least
max(4+d_x/2,1+5d_x/2); the second bound increases by1/2 if Gamma(x)!=0.
An active vertex with d_x=0 has pi(x)>=9/2. The single-class value2
bound f>=11 is recorded in that proof's section2 "Low N-lines"; its
f9 equality law is in section1 and the single-Z application in section4.
These allow balanced N. H section3.1 also proves the needed single-class
f<=10 values directly. A weaker f>=10 is used in I when another
nonzero class is present.

For B=(G,U={u,v}) define zeta_u=Z_u+D, zeta_v=Z_v+D and
beta=Z_u+Z_v+2D. Normalization removes D if an endpoint zeta vanishes;
Z attaches only in K. Set Gamma(x)=sum_B zeta_x(B)-1/2. A nonzero
class needs two N on distinct lines, one on each h side.

The exact mixed-free equations are:

    t_U^G=Omega_G+beta_(G,U),
    div(delta)_e=H0_e+sum_B beta_B(1[e=U]+1[e=V]/2+1[e=W]/2)-w_e,
    H0_uv=(B_uv+B_vu)/2,
    B_uv=-2 sum_(halves at u containing v) line_marginal,
    2 sum_B delta_B(1_W-1_V)=sum_B(Z_v-Z_u)(e_u-e_v),
    actual_star_sum_in_B+block_alternating_sum_in_B=beta_B.

Orient flow from V to W. Its divergence is -delta at V and +delta at W.
The loop identity is 2delta=L_W-L_V; a single Z and its sole L3 give
|delta|=|Z|/2. The group-array partition and block-alternating partition
are different. Their counts are not added without a checked disjointness
claim. Mixed columns vanish in all the equations used above.

The proved source identity at zero flow, when every nonzero zeta has
source U, is NT_U=w_U-1. Balanced classes elsewhere remain allowed.
The kernel move is (D,Zu,Zv,E)->(D+a,Zu-a,Zv-a,E+a). Z-free nonzero
flow costs at least6 loops. The fixed21-row dual is used only with
C=X=0 and its exact403-column scope, never as a mixed-column dual.

The modest private15 lemma used in F is proved from the private line equations: P15 gives positive stars and
A>=6. A three-glue private line forces all six active lines into one
split, whose stars have distinct block sources; otherwise there are
three single-glue generators, twelve positive stars, and at most two
stars per block. Ten private-split and216 generator controls are pinned.

Numbered references in Parts C through I are local to the indicated part.

## C. Nonmatching e5 and e6 exclusions



### 1. Inputs, budgets, and scope



Use the proved scale with center star masses1, signed correction weights
w, `sum w=6`, and `sum Omega=-2`. Write alpha_G=-Omega_G; then
`sum alpha=2`. Coefficients remain arbitrary signed rationals.

The proved canonical census excludes e<3 and e=4; e=3 is a perfect matching.
The complete possible width30 budgets under the current branch bounds are:

| e | R=30-e | Proved retained lower bound | Status |
|---|---|---|---|
| 3 | 27 | 27 | Matching branch; handled in the matching cases |
| 4 | 26 | — | Correction census excludes e=4 |
| 5 | 25 | 25 | Nonmatching boundary analyzed below |
| 6 | 24 | 24 | Nonmatching boundary analyzed below |
| >=7 | <=23 | 24 | Already excluded |

The proved inputs used here are canonical completeness, the correction
census, normalization, general V/h/array/flow equations, signed total mass,
center mass, f>=5 at inactive and f>=8 at active vertices (f>=9 when
Gamma!=0), source balance, the Z=0 nonzero-flow cost C>=6, and the
single-Z cost C>=5. These are general inputs; neither the matching
`Omega=-2e_M` reduction nor the pinned construction's 37-direction
obstruction is used.

Notation: `R=Pi+A+C+X`, `Pi=private+h`, `h=d+z`,
`private=stars+glues+N`, `m=|K|`, and
`q=#{G:Omega_G!=0}`. Nonmatching corrections have q>=2.

### 2. New charge: helpers can cancel at most one array each

**Lemma H.** Every normalized identity satisfies

    R >= ceil(15+3m/2) + 3q + C + X.                 (H)

Proof. Before helper cancellation, the q supported matching groups have
3q arrays with nonzero Omega. The exact array equations are
`t_U^G=Omega_G+beta_(G,U)`. Each D or Z belongs to one beta block and
therefore can help cancel at most one of these arrays. Multiple helpers
sharing a block still cancel at most that one array. At most h arrays
are canceled, so `A>=3q-h`. Extra nonzero arrays outside those groups
only strengthen this bound. Array supports are disjoint; zero-total
arrays are allowed to contain arbitrarily many canceling columns.

Meanwhile

    2(Pi-h)=sum f(u) >= 5(6-m)+8m,
    Pi >= ceil(15+3m/2)+h.

Adding the two bounds cancels the helper charge h and proves (H). No
sign assumption or addition of the block-star and array counts is made.
For q>=2, the m=2,3,4 lower bounds are respectively24,26,27 before C+X.
Thus the width30 nonmatching boundaries have **m<=2**. At m2, R<=25
also forces q2.

If m1 and z>=2, each Z is a different nonzero attachment class and block.
Each needs two N, and there is an actual star. Therefore

    Pi>=14+3z,    A>=3q-z,    C>=z,
    R>=14+3q+3z>=26.

For m1,z1, (H) and C5 give R>=17+6+5=28. Hence at either boundary,
m0/m1 has all zetas zero after normalization.

### 3. New effective-private support19 bound

**Lemma P19.** In an effective N=0 private layer with all center masses1,
private support P0<=19 implies

    H0_e <= 3/2 for every edge e.                   (P19)

H0 is not asserted to be nonnegative.

Proof. On a three-position private line with glue vector l and marginal
r=sum(l), the effective stars are `l-r*1` and the cost is
`F=g+2*star_count`. The complete signed spectrum is

    r!=0: F in {5,7,8,9};       r=0: F in {0,6,9}.

For one glue, F5 is automatic. For two nonzero glues a,b, the effective
stars are `(-b,-a,-a-b)`: F6 when a+b=0 and F8 otherwise. Three glues
give F5,7,9 at nonzero marginal, and F9 at zero marginal. At F5, every
nonzero glue has magnitude |r|: the forms are one glue r or three glues
`(-r,r,r)` up to permutation.

Center mass1 gives total marginal -1/2 at each vertex, so there are at
least six nonzero marginal lines. Total line cost is `2P0<=38`.

- With exactly six nonzero marginals, each vertex has one, of value
  -1/2. Zero marginals contribute nothing to H0. Every directed B row
  has two entries1 and the rest0; hence H0<=1.
- Eight nonzero marginals would cost40. Seven cost at least35; the
  allowed positive increments are2,3,4 and zero-line increments6,9.
  The only even total at most38 is **38: six F5 and one F8**, with no
  nonempty zero-marginal line.

In the second case one vertex u has two distinct private lines; the other
five have one each. The two u-lines belong to different private partitions.
A glue meets the opposite line in the same partition. Therefore at least
one F5 line at u meets only ordinary F5 lines of other centers: if F8 is
at u, its other line does; if F8 is at another center, at most one of the
two u-lines shares its partition. Those ordinary lines have marginal
-1/2 and every nonzero glue of magnitude1/2. The chosen u-line therefore
has marginal +/-1/2. Its companion is nonzero and their sum is -1/2,
forcing the two marginals to be **+1/2 and -1**.

Thus the duplicated center has directed entries
`B_uv=-2*sum(together marginals)` in `{0,-1,2,1}`. Every other center
has B entries0 or1. Since `H0_uv=(B_uv+B_vu)/2`, every edge is at most
3/2. This is an arbitrary-signed proof, not a coefficient grid.

When all zetas vanish, effective elimination creates at most one star
position per N column, so `P0<=stars+glues+N=Pi`. Its center masses
remain1. At zero flow and beta0, the general equation gives `H0=w`.
Consequently **P19 excludes any all-zero-zeta, zero-flow boundary with
Pi<=19 and a correction edge of weight2**.

### 4. Signed graph facts; no global q2 assumption

For e=5, the proved census supplies an edge of weight2. One can also
classify its signed degree-law graphs directly:

1. K_2 plus C_4, with K_2 weight2 and alternating cycle weights a,2-a,
   both nonzero.
2. The connected double-star, with four leaf edges weight2 and its
   central edge weight-2.

Reason: every vertex has nonzero graph degree. A leaf edge has weight2,
so its neighbor cannot have graph degree2. A connected six-vertex,
five-edge graph is a tree. The leaf-neighbor rule and degree sum10 leave
only degree sequence `(3,3,1,1,1,1)`; the two degree3 vertices join and
each has two leaves, giving the double-star weights. The other tree
degree sequences either have a degree2 leaf-neighbor or are the star,
whose center equation fails. In a disconnected graph, no three-vertex
component can fit e=5: its only degree-law possibility is a triangle, and
two triangles need six edges. The only remaining split is2+4. A connected
four-vertex four-edge component is C_4 or triangle-with-tail; the latter's
degree equations force a triangle edge zero. This leaves K_2+C_4.

An explicit signed control is `01=-2,02=03=14=15=2`. Every weighted
degree is2 and the signed edge sum is6. This control is **not a full PM2
identity**. It shows why e=5 must not globally be assumed to be a pair of
overlapping matchings. That pair description is used below only after
the retained budget forces q2.

For q2, two distinct matchings either share one edge (giving e=5) or are
disjoint (giving e=6), since alpha1+alpha2=2 and both alpha are nonzero.
For q3, each matching has a unique edge not in the other two: each other
matching can share at most one of its three edges. If all correction
weights equal1, each of these unique edges forces alpha_G=1, contradicting
`sum alpha=2`. This will exclude the low-budget e=6/q3 alternative.

### 5. Proposed five-edge improvement: R>=26

Assume e=5,R25, normalized. This width30 boundary is support-minimal by
the lower bound 30 bound. Lemma H excludes m>=3.

For m0/m1, section2 excludes any Z. All zetas vanish. Since A>=6,
`Pi<=19`. Nonzero flow would cost C6, giving at least27 at m0 or29
at m1. Thus flow is zero. P19 and `P0<=Pi` contradict the weight2
edge in the five-edge census. This covers arbitrary mixed coefficients
as well: the center, V/h and general F/source-balance rows have no X term.

For m2, (H) forces q2. If Z is present, C>=2: a sole loop forces all
Z into its block. A sole L3 leaves exactly one Z and contradicts C5;
a sole e forces equal endpoint Z, and the exact D/Zu/Zv/e move saves
support, contradicting the proved width30 minimum. Therefore (H)
gives R>=24+C>=26 when Z>0.

Let Z=0. Nonzero flow cannot fit by C6. If d0, all zetas are zero and
again A6,Pi<=19 contradict P19. If d>=2, each endpoint has at least
2d N and an actual star, while the four inactive centers cost at least10
private support. Thus

    Pi>=12+5d,    A>=6-d,    R>=18+4d>=26.

It remains one D, coefficient a, on source K. Here A>=5 and
`19<=Pi<=20`. Its group must be one of the two supported Omega groups;
otherwise all six supported arrays remain and its outside array adds
another, giving A>=7. Call that group G1, with alpha1, and the other G2,
with alpha2. The matchings share a correction edge e0 of weight2.

If K=e0, source balance gives `NT_K=1`. A nonzero separator costs at
least13 at one endpoint and at least8 at the other. Adding inactive cost20
and D gives Pi>=22, impossible.

If K is a unique G1 edge, its weight is alpha1. The same source balance
gives `NT_K=alpha1-1`; a nonzero value would again cost Pi>=22. Hence
alpha1=alpha2=1. If Pi19, a!=1/2 would make both Gamma values nonzero,
forcing active costs9+9 and Pi>=20. Thus a=1/2. If Pi20, the remaining
budget permits only A5; one of the six arrays must cancel, which requires
`2a=alpha1=1`. Again a=1/2.

Every inactive f is at most7: total line cost is at most38, the active
centers cost at least16, and the other three inactive centers cost at
least15. An inactive f<=7 has one nonzero marginal line of value-1/2;
any second such line costs at least5, and any nonempty zero line costs6.
Its B row therefore has two entries1 and the rest0.

The shared edge e0 is disjoint from K, so both its endpoints are inactive
and `H0_e0<=1`. The D helper is on G1 with e0 as a side; beta=1 supplies
only1/2 there. Zero-flow (F) consequently requires

    H0_e0 = w_e0-1/2 = 3/2,

a contradiction. All e=5,R25 cases, including the possible one-mixed
Pi19 boundary, are excluded.

### 6. The six-cycle private lemma and a signed erasure charge

**Lemma C18.** If an N=0 effective private layer has center masses1,
P0<=18, and H0 equal to the unit six-cycle, it is exactly the private
core with three glues -1/2 and twelve stars +1/2. Its twelve stars occupy
twelve distinct block-star charge blocks.

Proof. The private line count at P0<=18 leaves exactly six nonzero
marginals, one -1/2 at each center. The cycle edges of weight1 require
both directed B entries1, so each center's halfmates are its two cycle
neighbors. For cycle0123450 the selected halves are
`{i,i-1,i+1}`. Opposite centers i and i+3 select complementary halves.
Thus there are precisely three selected private partitions, with one
nonzero row and one nonzero column in each, both of marginal -1/2.

A nonempty zero-marginal line costs at least6 and has at least two glues.
It touches two opposite lines; at most one of these has nonzero marginal,
so it forces another nonempty zero-marginal line. Their cost at least12
added to the six nonzero lines' cost30 exceeds total cost36. Therefore
all zero-marginal lines are empty. Each selected private block contains
only its row/column intersection glue -1/2, yielding four +1/2 stars.
The fixed cycle core is checked directly against V, h, center masses,
all fifteen H rows, and the literal block-star ownership; the twelve
charged blocks are distinct. Relabeling covers every six-cycle. This
core is a forced private relaxation, not an asserted full identity.

**Lemma B (unchanged blocks).** With beta0, if the effective stars have b nonzero
block-star sums, then actual alternating support satisfies `A>=b-N_count`.
Each N contributes to one V-star position and hence at most one such
block. At least b-N_count block sums are unchanged and nonzero. The
literal B equation forces an alternating column in each; their ownership
sets are disjoint. Changed stars can have arbitrary signs. This count is
not added to any separate array lower bound.

### 7. Proposed six-edge improvement: R>=25

Assume e=6,R24. By (H), m>=3 is impossible.

For m0/m1, Z is excluded by section2. Nonzero flow would require C>=6, whereas Pi>=15 and A>=6 leave at most three loop terms; hence flow is zero. Since A>=6,
the effective private layer has center masses1 and P0<=Pi<=18. The
proved private18 lemma gives H0 in `{0,1/2,1}`. There are six nonzero
correction weights and their signed sum is6. Thus every weight is1.

At m1, q2 is forced by Pi>=17. At m0, q<=3; q3 is excluded by the
unique-edge argument in section4. Thus in both cases q2 consists of
disjoint matchings of weights1, and the graph is a six-cycle. Apply C18.

- At m0, actual and effective stars agree and force A>=12, contradicting
  R24 and private support15.
- At m1, total line cost is at most36; five inactive vertices cost25,
  so the active f<=11. Gamma=-1/2 requires at least one glue and actual
  center mass1 requires at least one star. Hence there are at most four
  N. Lemma B gives A>=12-4=8, while Pi>=17 permits A<=7. This treats
  arbitrary balanced coefficients and possible N collisions, without
  assuming actual stars are positive.

For m2, (H) forces q2 and C=X0. Hence Z=0. With d0, A6 plus Pi>=19
already gives25. With d>=2 the `18+4d` bound above gives at least26.
With d1, exactly Pi19,A5 remains. Its D group must be supported and its
array must cancel. The Pi19 cost forces a=1/2 and glue-free active
endpoints, with simple f5 inactive vertices. Therefore alpha1=2a=1 and
alpha2=1. e=6 means the two matchings are disjoint. At an active endpoint u
of the D source, its G2 edge e lies outside G1, so receives no helper.
Its other endpoint is inactive. Thus `H0_e<=1/2` but zero-flow (F)
requires `H0_e=w_e=1`, a contradiction. All e=6,R24 cases are excluded.

## D. Matching rigidity and the m4,m5,m6 exclusions



### 1. Helper--array count and stronger matching rigidity

Let h=d+z count helper directions, q the support of Omega, and m the
number of N-attachment vertices. Keep all notation and normalization
from the lower bound 30 proof. Exactly3q group arrays have nonzero
Omega. At most h blocks can have nonzero beta: each D or Z belongs
to just one block. A supported-group array can vanish only in such a
block. Every other supported-group array has a nonzero total and needs
an alternating column. Therefore

    A>=3q-h.

Meanwhile 2Pi=sum f+2h>=5(6-m)+8m+2h, giving

    Pi>=15+ceil(3m/2)+h,
    R>=15+ceil(3m/2)+3q+C+X.                              (HA)

The argument is signed, uses the group-array partition only, and does
not double-count a helper shared with an N class. A helper may fail to
cancel any array, but cannot cancel two. A block containing several
helpers still cancels at most one array.

For matching corrections, a nonzero Omega+2e_M relation has at least
six matching-incidence terms, by the proved signed kernel lemma.
Thus any Omega other than -2e_M has q>=5. Formula (HA) then forces
R>=30, even at m0. Consequently

    matching R<=29 implies Omega=-2e_M.                   (O29)

This extends the scope of the prior R<=26 rigidity without assuming
that the old statement already covered R27. For the remainder assume
matching R27 and use (O29).

The proved matching lower27 makes every support-decreasing normalized
kernel move impossible at R27. Thus C1 remains impossible when Z is
present; with Z=0 the S rows also rule out one nonzero loop.

### 2. Two new small facts

**No f10 at a single value1 helper class.** Suppose the vertex has
one nonzero zeta class of value1, no other nonzero helper class, and
Gamma=1/2. Then f10 is impossible. Three N in the main class use
three lines and cost12. Any extra balanced class adds at least two N;
four N, one actual star and one glue cost11. Thus f<=10 leaves the
main pair N=-1,-1 on distinct lines and no other N. Additional nonempty
lines cannot fit. The only exact cost10 pairs are F4+F6 or F5+F5.
Their total marginals are respectively0 or one of1,3/2,2 by the full
collision-safe F4/F5/F6 lemma; none equals1/2.

**Two unit D columns cannot fit Pi<=26 and A<=2.** Suppose the only
helpers are D=1 in (M,U) and (M,V), with U,V disjoint matching edges,
and K=U union V. Assume C=0 and Pi<=26. The third matching array
has total -2, whereas the other two and all outside-M arrays have
total0. An outside-M alternating array would need two columns in
addition to the nonzero M-array, so A<=2 makes every alternating
column a group-M column.

The block-alternating columns in each of the three group-M loop blocks
are disjoint from all twelve group-M alternating columns. This is an
explicit incidence fact, checked on all45 block/group combinations;
it is not an identification of the two alternating partitions. Hence
the two D blocks have no block-alternating coefficient. Identity (B)
requires actual star sum2 in each of them, a combined sum4.

But each active endpoint has Gamma1/2 and f>=9, and each inactive
vertex f>=5. Including d2, Pi>=25. The line-cost budget up to Pi26
has only two units of slack above active (9,9,9,9) and inactive (5,5).
An active f10 is impossible by the preceding fact. If an inactive
vertex had f7, all other vertices would be at their minimum: every
active f9 glue is +/-1/2, and the other inactive f5 glues are in
{0,+/-1/2}. Every glue at the f7 vertex crosses one of those lines,
so it too is on this grid. There is just one nonempty line at that
vertex; its sum is -1/2. On the forced grid such a line cannot have
cost7 (the F7 two-star form requires a nonstandard entry). Thus the
inactive vertices both have f5. Every active f is at most11.

Each active endpoint has two main class lines, costing at least4 each.
A third main N line, or any other nonempty private line, would cost
at least12. Therefore all actual star support at an active vertex lies
in those two main lines. Their halves contain its D source pair, so
their star source pairs contain its mate and exclude the center. Such
a source is neither U nor V. This is checked for every V position on
every possible main line, not just the N position.

Only the two inactive centers can therefore contribute stars in the
two D blocks. Their f5 patterns have nonnegative stars, and each has
total actual star mass1. Their combined contribution to any subset of
blocks is at most2, contradicting the required4. Actual stars at a
possible active f11 vertex need not be nonnegative; they are excluded
from the two blocks by their source incidences. This proves the claim.

### 3. Matching R27 excludes m>=5

At m6, Pi>=27. Equality is the stated private equality case with
Z=0 and disjoint D=1/2; all three M-arrays remain nonzero, so A>=3.
Larger Pi already exceeds R27.

At m5, Pi25 equality similarly requires A>=3. Pi27 leaves no coverage
direction. At Pi26 only one direction remains outside Pi. A Z
would force C>=2, so Z=0. The M-edge containing the inactive vertex
cannot carry D, and its unchanged array has total -2. Hence A>=1,
forcing A1,C=X=0. Canceling two M-arrays requires two D=1 on two M-edges
inside K. No extra D can occur: an outside-M block would create another
array, and the remaining M-edge has an inactive endpoint. Their four
endpoints have f>=9; the fifth active zero-zeta vertex also has f>=9;
the inactive vertex has f>=5. Including d2 gives
Pi>=(5*9+5)/2+2=27, contradiction.

### 4. Matching R27, m4, nonzero helpers

The **nonzero-helper hypothesis is essential** in this section. The
all-zero-zeta m4 case is covered by Part F.
Now Pi>=23. Pi27 leaves no coverage direction.

**Pi23.** Equality has Z=0 and disjoint D=1/2. A>=3 leaves C<=1,
so C0; X may be0 or1. There are one or two D. One contradicts
source balance at its glue-free endpoints, whether its block is in M
or outside. Two outside-M D would give A>=5, exceeding the budget4.
Thus with two D at least one is in M, on source U. Formula (F) there
requires H0_U=2-1-eta, where eta is0 or1/2 from the other helper.
This is positive, whereas the glue-free endpoints give0. These
contradictions permit A4 or the possible mixed column; neither case
is silently discarded.

**Pi24 with Z.** C>=2 leaves A<=1. Here h<=3. If A0, all three
M-arrays must cancel, so K meets every M-edge and consists of one
whole edge and two crossing endpoints. Exactly three helpers are
available, one per canceled array. The two crossing Z must be2;
the third helper is unit D or Z of coefficient 2 on the whole edge. Every active Gamma
is nonzero, so f>=9 at all four vertices and Pi>=5+18+3=26.

For A1,C2, exactly two M-arrays must cancel. A zero-beta block outside
M would need at least two helpers in addition to at least one in each
canceled block, exceeding h3. A nonzero-beta block outside M would
need another alternating. Thus all helpers lie in M-blocks. C2 permits at most two Z blocks. If the whole M-edge U is canceled
by unit D, a crossing M-edge e_2 by Z_x=2, and the third crossing M-edge e_3 carries Z_y!=2, the active costs are at least9,9,11,8. With
inactive cost10 and h3, Pi>=ceil(47/2)+3=27>24. Otherwise, a Z in the third M-block forces both canceled blocks to
contain Z, and a third Z block would need C>=3. A D in the third,
uncanceled M-block is possible only when the two canceled arrays are
the crossing ones; then those two single Z are2 and cost f>=11
each; the two D endpoints cost at least8 each. Including h3 and
inactive cost10 gives Pi>=27. Thus after this exclusion all helpers
lie in the two canceled blocks.

Here a single Z globally would cost C>=5, so z>=2. If h2, the two
helpers are Z of coefficient 2 at distinct active endpoints. The other
two active endpoints have zero zeta, giving costs at least11,11,9,9
and Pi>=27. If h3, one possibility is a Z pair totaling2 in
one canceled block and unit D in the other. The one-block lemma forces
that pair to be1,1. All four Gamma are nonzero,
so Pi>=26. Otherwise a single Z of coefficient 2 is in one canceled block and a
two-helper combination in the other. The Z of coefficient 2 endpoint costs11; an
active vertex outside the other helper source costs at least9; the
two remaining vertices cost at least8 each. Thus Pi>=5+18+3=26.
No Pi24 placement survives.

**Pi24, Z=0.** A>=1 and the budget gives C<=2. If C2 then A1,
but A1 requires two unit D with Pi>=25; so C0. One D has source balance
and costs Pi>=26: if a!=1/2 the endpoint costs are at least13,9,
and if a=1/2 they are at least14,8; the other active costs are9,9
and inactive costs5,5. Three or more D cost at leastPi26 by the
proved D-degree bound restated in Part B. Two D with the same source or overlapping
sources also cost at least26 or25 respectively; only disjoint sources
can fit Pi24.

At most one of those D blocks can be outside M. If one is outside,
A<=3 requires cancellation by unit D in the M block. The other D must
be1/2 to fit Pi24; otherwise all four active Gamma are nonzero and
Pi>=25. The M endpoints have f9, so H0_U=-1. Formula (F) requires
H0_U=0 or-1/2 according as U is absent from, or a side of, the other
helper group. Contradiction.

If both D are in M, either both coefficients are1/2, or exactly one
differs from1/2. If both differ, Pi>=25. With both1/2, each endpoint
has f<=10; every nonzero-marginal line must be a together line because
a separator would cost at least13. Thus H0_U=0, whereas (F) needs
1/2. With just one different coefficient a, tight cost forces its
endpoints f9, the other endpoints f8 and inactive f5. The F5 marginal
equation a-1/2 in {a,a/2} forces a=1. Now H0_U=-1 whereas (F)
needs -1/2. This excludes Pi24, including X1 when A2.

**Pi25 with Z.** Only A0,C2 can fit. Now h<=4. Again all three
M-arrays must cancel and K contains one M-edge and two crossing
endpoints. Helpers outside those M-blocks cannot occur: they would
need at least two helpers for beta0, beyond the three required
cancellations. The crossing Z are2. At h3 the whole-edge helper is
unit D or Z of coefficient 2, giving Pi>=26 as above. At h4 the two whole-edge endpoints
cost at least8 each and the crossing endpoints at least11 each;
including h4 gives Pi>=5+19+4=28. Contradiction.

**Pi25, Z=0.** A>=1 makes C<=1, hence C0. If A1, two unit D in M
are forced. The two-D lemma in section2 excludes A1 and also A2
with those coefficients. For A2, at least one M array cancels, so
there is a unit D in (M,U). One D costs Pi>=26 and three cost Pi>=26.
For two D, a common source costs Pi>=26. If their different sources
overlap, the second source is not an M-edge, since the first is U in M.
Its D therefore has an outside-M group and a nonzero outside array.
The two unchanged M-arrays plus that outside array force A>=3>2.
Disjoint sources remain. A second outside-M D would give
two unchanged M-arrays plus its own, exceeding A2. Thus the second
D is in (M,V), coefficient b.

If b is neither1/2 nor1, all four Gamma are nonzero and Pi25 forces
all active f9; the F5 equation would force b=1, contradiction. If
b=1/2, every endpoint has f<=11, below the separator cost13, so
H0_U=-1 still; (F) needs -1/2. If b=1, section2 excludes it. Thus
Pi25 is impossible.

**Pi26.** Only A1,C=X=0 can cover. Z is impossible. Canceling two
M-arrays forces precisely two unit D in M: any additional D has a nonzero
array outside M, or would use an inactive endpoint in the remaining
M-edge. The two-D lemma excludes Pi26,A1. This finishes every m4
case with nonzero helpers.

## E. Private support at least25 under H0=2M



### Statement and stated definitions

Every vertex has total center mass one. For each private 3|3 partition
sigma and vertex u, let a_(sigma,u) be its center mass on that line.
The directed marginal matrix is

    B_uv = sum_(sigma: u,v lie in the same half) a_(sigma,u).

Assume H_uv=(B_uv+B_vu)/2 equals 2 on the edges of a perfect matching M
and zero on every other pair. Then the private support satisfies P0>=25.

The stated private line spectrum is as follows. A nonzero-marginal
line has cost in {5,7,8,9}; a nonempty zero-marginal line has cost 6 or 9.
The sum of line costs is 2P0. A simple vertex has exactly one nonzero
center-mass line; its mass is one and its B row has exactly two entries
one, all other entries zero. These facts are proved in Appendix A of
the proved lower-27 proof and the uniform matching reduction.

### Proof

Suppose P0<=24. No matching edge can have two simple endpoints: its H
value would belong to {0,1/2,1}, rather than equal two. At least three
vertices are therefore nonsimple, and at least nine lines have nonzero
marginal. Ten such lines already cost at least 50>2P0. Thus there are
exactly nine: precisely three vertices are simple, and the other three
have exactly two nonzero-marginal lines each. Every matching edge joins
one simple vertex to one two-line vertex.

Let v be a two-line vertex and s its simple matching partner. The two
line masses at v sum to one. Since B_sv is zero or one, the equation
B_vs+B_sv=4 forces B_vs to be 4 or 3. The partner s must consequently
belong to exactly one of v's two half triples: belonging to both gives
B_vs=1, and belonging to neither gives zero. Orient the two masses as
p and 1-p, where p is 3 or 4. For any other vertex y, its membership in
those two halves gives

    B_vy in {0, 1, p, 1-p}.

In particular, **no entry in the row of any two-line vertex is -1**.
No simple vertex has an entry -1 either.

Now fix a simple vertex s. Its unique half has two other vertices, so
its row has two entries B_sy=1. At most one of those vertices is s's
matching partner. For the other one, H_sy=0 requires B_ys=-1. The
preceding paragraph excludes this for every possible y. Contradiction.

Therefore P0 cannot be at most 24, and P0>=25.

The proof never assumes that a nonzero center mass is positive. It also
allows nonempty zero-marginal lines: they affect support cost but not B,
and cannot change the nine-line argument.

## F. All normalized zetas zero, every attachment count



Use the stated notation R=Pi+A+C+X, effective private support P0<=Pi,
and the rigidity (O29) of Part D section1. When all
zetas vanish, normalized D=Z=0, beta=0, and A>=3. All effective center
masses equal one. With Z=0 every nonzero-flow block costs at least two
loop columns, and a nonzero flow needs at least three blocks.

### 1. Zero flow

Part E gives H0=2M => P0>=25. Thus R>=P0+A>=28. This excludes zero flow for
every number of N attachment vertices, including the previously retained
Pi=P0=24,A3 boundary.

### 2. A three-edge flow forces the six-cycle core

Suppose P0<=18 and a nonzero flow has at most three supported blocks.
The stated private-18 lemma gives H0 entries in {0,1/2,1}. Moreover Z=0 makes sum_B delta_B*c_B=0 by projection. Summing
edge divergences over pairs containing a coordinate x gives
sum_(e containing x) div(delta)_e=sum_B delta_B*c_B(x)=0. With
beta0, (F) then gives sum_(e containing x)H0_e=sum_(e containing x)w_e=2.
Thus H0 has coordinate degree two and its edge weights sum to six. The flow divergence is H0-2M.

Each of the three matching pair-nodes has strictly negative demand.
Their total negative demand is 6-sum_(e in M) H0_e>=3. All other
demands are nonnegative and at most one, so at least three other
pair-nodes have positive demand. Three flow edges can touch at most
six demanded nodes. Therefore there are exactly six, and the flow
consists of three disjoint edges, each joining a positive demand to a
negative demand. This uses signed flows correctly: every connected
component has zero total divergence, regardless of edge orientation.

There are exactly three positive demands, each at most one. Their sum
is at least three, so all equal one. It follows that H0 equals one on
all three matching edges and on exactly three further edges, and is
zero elsewhere. Its support is a simple 2-regular graph on six vertices.
The only possibilities are a six-cycle or two triangles. The latter
contains no perfect matching, whereas the support contains M. Thus H0
is a six-cycle and M is one of its two alternating perfect matchings.

The effective private layer is then the stated fixed six-cycle core:
three glues of coefficient -1/2 and twelve stars of coefficient +1/2,
in twelve distinct star blocks. For completeness, the private-18 lemma
forces exactly one nonzero-marginal line per vertex, of sum -1/2.
Its halfmates must be the two cycle neighbors. Opposite cycle vertices
therefore give the two nonzero-marginal lines in each of three splits.
The available excess over their baseline cost 30 is at most six, so
there is at most one nonempty zero-marginal line. Such a line requires
two nonzero crossing cells, but the opposite half has only one other
nonempty line; no second zero-marginal line can fit. This is impossible.
With all other lines empty, each used split has just one glue between
its opposite cycle vertices. Hence P0=15 and the stated twelve stars
are forced. Their distinct literal block incidences are already pinned
and are independently reconstructed by the literal-incidence checker.

### 3. No N directions

Here Pi=P0=P. The stated private support gap is P=15 or P>=18.
At P15 the proved private15 lemma gives A>=6. Nonzero flow gives
C>=6, so R27 forces A6,C6,X0 and at most three flow blocks. Section2
then gives the six-cycle core. With N0 its twelve blocks have positive
actual star sums and beta0, so (B) gives A>=12>6, contradiction. This
uses only the modest private15 incidence lemma restated in B above.
For P>=18, R=27, A>=3 and C>=6 force P=18,A3,C6,X0. There are at
most three nonzero-flow blocks, so section 2 forces P=15, contradiction.

### 4. At least one N direction

The stated charge gives Pi>=17. From R27,A>=3,C>=6 we obtain
Pi<=18 and C<=7. Thus the nonzero flow again has exactly three blocks.
Section 2 gives the effective six-cycle core P0=15.

The charge Pi>=15+2m confines N to a single attachment vertex u. The
five inactive vertices cost at least five each, so f(u)<=2Pi-25<=11.
Gamma(u)=-1/2 requires at least one glue, and the actual center row
requires at least one actual star. Consequently

    1+2+2N <= f(u) <= 11,  hence N<=4.

Each N direction changes exactly one effective star cell, and every
star cell belongs to one block. At most four of the twelve positive
core blocks can change. At least eight therefore retain their original
positive actual star sum 1/2. Since beta=0, identity (B) requires at
least one alternating column in each of these eight disjoint blocks.
Thus A>=8. The budget instead gives A<=27-17-6=4, contradiction.

No sign condition is imposed on stars in changed blocks. N collisions
can only reduce the number of affected blocks. The argument does not
add the block-alternating count to the group-array count.

## G. One attachment vertex and useful m2/m3 reductions



The proof uses the helper-array bound and matching rigidity R<=29 already
accepted in Part D, together with the lower bound 30 general equations.
Write K={u}. Normalization excludes D, since a surviving D requires two
nonzero attachment endpoints. Every Z is at u, in its own block/class.
The sole matching block available at u is (M,{u,u^ast}). Other Z blocks lie
outside M. Hence A>=2 always, and for z=2 one has A>=3. If z=2 and
A=3, exactly one Z block is in M and its coefficient is 2.

### 1. A single Z requires at least six loop directions

**Lemma L1+.** In any normalized identity with exactly one nonzero Z,
the number C of loop directions is at least six. D may be arbitrary.

Let the Z have coefficient z!=0 at u in B=(U={u,t},V,W). Divide the
projected flow equation by z. Its right side is rho=e_t-e_u. A block
flow contributes q_B c_B, where c_B=1_W-1_V and q_B=2delta_B/z.
The vector c_B vanishes exactly on its source U.

If the Z block contains one loop direction, its S rows force that loop
to be an L3, and the proved exact scale gives q_B=+1 or -1. The residual
rho-q_B c_B has six nonzero coordinates, each +1 or -1. One other block
cannot supply it. If two other blocks supplied it, their four-element
supports would have to cover all six coordinates, with intersection
size two and two exclusive coordinates apiece. The exclusive coordinates
force both flow coefficients to have absolute value one. At a shared
coordinate their sum is then 0 or +/-2, never +/-1. Thus at least three
further blocks are necessary. All are Z-free and need at least two loop
directions each, giving C>=1+3*2=7.

If the Z block has at least two loop directions, its residual has support
two when q_B=0 and support six when q_B!=0. It cannot equal one nonzero
four-supported block vector. At least two other nonzero-flow blocks are
therefore required. They are Z-free, so C>=2+2*2=6. This proves L1+.

This is a signed argument. D never appears in the projected equation or
the relevant S rows. The proof does not classify all three-block circuits
and does not use the prior component separate FS(a) assertion.

### 2. Two Z blocks with at most three loop directions

Suppose the two Z coefficients are a,b at u, on sources {u,t_1},{u,t_2}.
If C<=3, every flow block must be one of these two: an additional Z-free
block costs two loops besides the required one in each Z block. The
projected equation at coordinate u therefore gives a+b=0, since both
block vectors vanish at u. In particular Gamma(u)=-1/2.

If the sources coincide, the projected right side is zero. The two
distinct block vectors are nonparallel, so both flows vanish. Each
single-Z block then needs at least two loops: one L3 alone has nonzero
flow by the exact scale used in section 1. Thus C>=4, contradiction.
Consequently, whenever C<=3 the sources are different and b=-a.

The two different-source N classes have at most one common private
line. Their common half, when used, is exactly {u,t_1,t_2}. Each nonzero
class needs at least two N on distinct lines. Therefore four N need at
least three lines. Five N need at least four lines: they must be split
as three in one class and two in the other, since any additional balanced
class needs at least two N. These are the proved class-incidence facts.

If one Z block is in M, relabel t_1=u^ast, with coefficient a. At both source
edges {u,u^ast} and {u,t_2}, the actual flow divergence is zero: neither edge
can be a side of either block because it contains u. Neither helper
contributes to the other one's source edge for the same reason. Equation
(F) consequently gives the exact pair of identities

    H0_(u,u^ast)=2-a,       H0_(u,t_2)=a.                 (E)

### 3. Low-cost common line with two opposite N coefficients

The following collision-aware local fact will be used at Pi22. Suppose
one line contains precisely two N, of coefficients n and -n, n!=0, and
has total cost at most seven. Then its possibilities are:

    F4: no glue and no actual star; the two N occupy the same cell;
    F6: marginal zero;
    F7: marginal +n or -n.

Proof. Write v for the N sum-vector, whose entries sum to zero; the two
N may collide. With glue vector l and s=sum(l), the actual star vector
is l-s*1-v. After the four units charged to the two N, at most three
remain for glues plus twice the actual-star count.

With no actual stars, summing the star equation gives s=0, hence l=v.
This gives F4 if the N collide, and F6 otherwise. With an actual star,
there is exactly one star and at most one glue. A nonzero glue and the
two opposite N can leave exactly one star only when the N occupy the
two nonglue positions, the glue coefficient is +n or -n, and the actual
star coefficient is twice its negative. This is F7 with s=+/-n. No other
case fits the cost. The accompanying checker verifies these affine
equations after the legitimate homogeneous normalization n=1; it does
not impose a coefficient grid.

For a line with one of the main N, the proved full F4/F5/F6 lemma gives
F4 glue-free, F5 marginal -n or -n/2, and F6 marginal zero. This includes
the collision-safe scope of the proved lemma.

### 4. Complete helper and support budgets

If z>=3, each helper needs two N. The five inactive vertices contribute
at least 25 to sum f, and u has at least one actual star. Thus
Pi>=14+3z. Each Z block needs a loop, C>=z, and A>=2, giving R>=28.
It remains z=1 or z=2.

#### 4.1 One Z

Here Pi>=18 and C>=6 by L1+. An outside-M Z needs A>=4, already giving
R>=28. If A=2, the M array cancels, so Z=2 and the proved single-class
bound gives Pi>=19. The only width-27 budgets are therefore

    (Pi,A,C,X)=(19,2,6,0), with Z=2;
    (Pi,A,C,X)=(18,3,6,0).

The first has active f11 and inactive f5. The proved single-Z=2 f11
argument excludes it: crossed f5 lines force the half-grid; there are
precisely two main N=-2, and their possible low-cost line marginals
cannot sum to Gamma=3/2. The second has active f9 and inactive f5.
The proved f9 argument forces Z=1 and a positive dense line whose
crossing lines give unequal marginal totals on the two halves of its
private split. These are the two single-Z arguments in section4
of the lower-bound-30 component above: the paragraph
starting "If z=1" and the following paragraph starting "If A>=3".
They use only active f11 or f9, inactive f5, the V/h/center rows and
crossing-glue magnitudes. C,X and flow do not enter; the arguments
therefore apply unchanged to these C6 budgets.

#### 4.2 Two Z: exhaustive tuples

Now Pi>=20, A>=3 and C>=2. The two Z are at the same endpoint and
therefore occupy distinct blocks, each of which needs a loop.
Distributing the two remaining directions gives
exactly these ten tuples:

| Pi | A | C | X | Exclusion below |
|---:|---:|---:|---:|---|
| 20 | 3 | 2 | 2 | I |
| 20 | 3 | 3 | 1 | I |
| 20 | 3 | 4 | 0 | I |
| 20 | 4 | 2 | 1 | II |
| 20 | 5 | 2 | 0 | II |
| 20 | 4 | 3 | 0 | II |
| 21 | 3 | 3 | 0 | III |
| 21 | 3 | 2 | 1 | III |
| 21 | 4 | 2 | 0 | III |
| 22 | 3 | 2 | 0 | IV |

**I. Pi20,A3, arbitrary allowed C/X.** The M Z equals 2. At least four N
and one actual star give active cost at least 10.
The total line cost is 36; inactive costs are 5 or at least 7. Thus
every inactive vertex has f5 and the active cost is exactly 11. Its
odd cost requires a glue, making the costs tight: exactly four N, one
star and one glue. The glue is +/-1/2. If the other Z is b,
Gamma=2+b-1/2 is +/-1/2, so b=-2 or -1. Every N coefficient is now an
integer. One half-integral glue creates two nonzero half-integral
effective star entries. Integer N cannot remove either, contradicting
the single actual star. This is the proved half-integrality argument
and imposes no condition on C or X.

**II. Pi20,A>=4,C<=3.** Section 2 forces different sources and b=-a.
The four required N use at least three lines, each costing at least
four. That contradicts active f<=11. Equal sources were already ruled
out by their zero-flow C>=4 cost.

**III. Pi21,C<=3.** Again the sources differ and b=-a. At least three
N-lines are needed; Gamma=-1/2 rules out all three being F4. Thus f(u)>=13.
The total line budget is 38, so active f13 and inactive f5 are forced.
There are exactly four N: five would occupy four lines costing at least
16, and six plus an actual star and glue already cost at least 15.

The two classes therefore use exactly three lines, with their overlap
containing opposite N. The other two lines cost at least four each, so
the common line costs at most five. Section 3 makes it the glue-free F4
collision. The other two lines must be F4 and F5. Hence the active
vertex has a unique nonzero marginal line, of marginal -1/2; this line
is not the common class line. Every B row has only entries zero or one,
so both H0 values in (E) lie in [0,1]. A<=4 ensures one Z block is in M,
since two outside blocks would need A>=5. Applying both equations (E)
forces a=1 and both H0 values to equal one. The active half must then
contain both u^ast and t_2. It is exactly the common class half {u,u^ast,t_2},
contradicting that the common line is glue-free.

**IV. Pi22,A3,C2.** The M Z is a=2, the other is b=-2, and the sources
differ. The total line budget is 40. As above active f>=13, so active
f<=15 and every inactive f<=7. An inactive f<=7 has exactly one
nonzero-marginal line of sum -1/2. Its B row therefore has only zeros
and ones, even if its individual glues are not on the half-grid.

There are at least four N and at most six, since seven N plus a glue
and an actual star cost at least 17. Five N would use four lines and
cost at least 16, so N is four or six.

If N=6, its twelve units leave only one glue and one actual star within
f<=15. The active vertex has a single nonzero-marginal line of sum -1/2,
so its B row is also zero/one. Equation (E) at the non-M source would
require H0=2, whereas H0<=1. Contradiction. This covers the extra balanced
N class that must not be silently excluded from the f15 boundary.

If N=4, there are exactly three N-lines; a fourth nonempty line would
cost at least five beyond their baseline 12. Two N of opposite values
2,-2 occupy the common line, and the other two lines each have one N.
The common line has cost at most seven. If it is F7, its marginal is
+/-2, and the other lines must both be F4; they cannot sum to -1/2.
If it is F6, its marginal is zero and the other two lines cost at most
nine, leaving F4+F4 or F4+F5. F4+F4 gives total marginal0, and F4+F5 gives a marginal in
{+/-1,+/-2}. Neither equals -1/2. If the common line is F4, the two remaining lines
have combined cost at most eleven. Apart from F4+F7, every possible
pair drawn from F4/F5/F6 has an integer marginal sum, not -1/2.
Thus only F4+F7 survives the local sum condition, and the F7 is the
unique nonzero-marginal line at u. Once again its B row is zero/one,
and (E) requires the impossible H0_(u,t_2)=2. This finishes IV.

### 5. Immediate reductions of the remaining m2/m3 cases

**Exactly one Z is impossible at m2 or m3, even with D present.** At
m2, Pi>=20: if d>=1 this follows from Pi>=18+d+z; if d0, the active
vertex without a helper has f>=9 and the other has f>=8, giving
Pi>=ceil((20+9+8)/2)+1=20. Also A>=2. If K is an M-edge, the other
two M-arrays are untouched. If K is not an M-edge, every D is on source
K outside M, and the single Z can cancel at most one M-array. Lemma
L1+ therefore gives R>=20+2+6=28.

At m3, Pi>=22 with one Z. If d>=1 the generic bound gives
Pi>=20+d+z>=22. If d0, two active vertices have f>=9, the Z endpoint
has f>=8, and the inactive cost is at least 15, so
Pi>=ceil((15+9+9+8)/2)+1=22. Then Pi+C>=28 already, by L1+.

**With Z=0, nonzero flow is impossible at m2 or m3 and R27.** At m3,
at most one M-edge lies inside K, so at least two unchanged M-arrays
need A>=2; Pi>=21 and C>=6 already give R>=29. At m2, A>=2 and
Pi>=19. If Pi>=20, R>=28. In Pi19 equality, either all helpers vanish
(A>=3) or the proved equality classification gives one D=1/2, which
cannot cancel any M-array (again A>=3). Thus R>=19+3+6=28.

These are preliminary reductions, not complete m2/m3 exclusions.
The remaining possibilities there have either Z=0 and zero flow, or
at least two Z. Arbitrary idle loops are not excluded unless their
support budget or the S equations exclude them separately.

## H. Two attachment vertices: complete two-attachment proof



### 1. Inputs and notation

The preceding lower-bound-30 component supplies the input. The helper-array argument is restated below before its use.

Use scale3 from the proved proof: every center has actual star mass1.

    R = Pi + A + C + X,
    Pi = stars + glues + N + D + Z,     C = L3 + E,
    h = d+z,                          2Pi = sum f(u)+2h.

For a private line, F=g+2s+2n; f(u) is the sum over its lines. Inactive
vertices have f>=5 and first strict excess>=2. Active vertices have f>=8,
or f>=9 if Gamma!=0. For K={u,v}, all D have source K. Normalization
removes a D if either of its endpoint zetas is zero.

For block B=(G,U), with U={a,b},

    zeta_a=D+Z_a,  zeta_b=D+Z_b,  beta=2D+Z_a+Z_b,
    Gamma(a)=Psi(a)-1/2,        Psi=sum zeta_a.

A nonzero class (G,U,a) has four possible N on four distinct lines whose
halves contain U. Its two h-side sums are -zeta. Precisely two N therefore
have coefficients -zeta,-zeta. A balanced class needs at least two N.

The proved equations used are:

    t_U^G=Omega_G+beta_(G,U),
    div(delta)_e=H0_e+sum_B beta_B*(1[e=U]+1[e=V]/2+1[e=W]/2)-w_e,  (F)
    H0_ab=(B_ab+B_ba)/2,
    B_ab=-2*sum_(halves at a containing b) line_marginal,
    2*sum_B delta_B*c_B=sum_B (Z_b-Z_a)*(e_a-e_b),                  (P)
    c_B=1_W-1_V,
    actual_star_sum_in_B+block_alternating_sum_in_B=beta_B.        (B)

Orient every block V->W: its edge divergence is -delta at V, +delta at W,
and zero at U. Thus |delta|, not |2delta|, is the side-edge magnitude.
The literal identity is 2delta=L_W-L_V. A sole Z and sole compatible L3
give L3=-Z and |delta|=|Z|/2. The checker pins all45 orientations and
literal scales, all360 such Z/L3 controls, and all15 effective F rows.

Mixed columns vanish in V, h, (F), (P), (B) and center mass. The fixed
21-row dual is used **only at C=X=0**. Its complete scope is all270
private columns, all120 N attached in K, all12 group-M alternatings, and
the one D(M,K): 403 columns. The dual combines homogeneous,
correction-free kernel rows of the reduced model. The fifteen correction
columns and the target do not enter these rows. Its equation is

    -D(M,K)+t_K^M/2=0.                                         (D21)

For a single D in M, t_K^M=-2+2D, so (D21) gives -1=0.
Relabeling covers all matchings and matching source pairs.

Other proved inputs: the signed matching-incidence kernel needs at least
six supported terms; matching retained support is at least27; one global
Z costs C>=5; Z=0 with nonzero flow costs C>=6; each Z block needs a loop;
a Z-free nonzero-flow block needs at least two loops. At R27 a
support-decreasing exact kernel move contradicts matching lower27:

    (D,Za,Zb,E) -> (D+t,Za-t,Zb-t,E+t).                         (K)

In particular C1 is impossible. With Z=0 this also follows directly from
S rows. When all nonzero zetas have source K and flow is zero, the proved
source balance, allowing balanced N elsewhere, is

    NT_K=sum_(separating lines at u,v) marginal=w_K-1 in {1,-1}. (T)

### 2. Matching rigidity at R27

Let q be the support of Omega. Its 3q supported-group arrays each have
nonzero Omega. A helper can change only its own array, so

    A>=3q-h,  Pi>=ceil(15+3m/2)+h,
    R>=ceil(15+3m/2)+3q+C+X.

Unless Omega=-2e_M, the signed relation Omega+2e_M is nonzero, has support
at least6, and gives q>=5. The last inequality then gives R>=30 even at
m0. Thus R27 implies **Omega=-2e_M**. Every M-array has total beta-2;
every other array has total beta. Nonzero array totals need separate
alternating columns. A zero-total array with nonzero alternating support
needs at least two columns.

Because some M-edge avoids K, at least one M-array remains unchanged:
**A>=1** for the entire m2 argument.

### 3. Collision-safe local facts

#### 3.1 Low-cost lines and one-class vertices

The proved full line lemma, including N from other classes, says: F4
N-lines are glue-free; an F5 line containing specified nonzero N=n has
marginal -n or -n/2; an F6 N-line has marginal0. Any F5 line with nonzero
marginal r has each nonzero glue of magnitude |r|. This last statement
also holds with no N: its form is one glue or three signed equal-magnitude
glues. The checker independently verifies the small exact scalar masks
with one or two N and arbitrary rational second coefficient; no grid is
imposed. For the marginal bound alone, any three-N F6 line would
have g=s=0 because 2n=6, and would therefore have marginal0. No
per-position incidence assertion is needed for this statement.

Suppose a vertex has just one nonzero class, value a, and f<=10.
If it has four or more N, f<=10 forces exactly four N, one actual star
and no glue, so Gamma0 and a=1/2. Three main-class N use three distinct
lines and cost at least12; additional balanced classes require two N.
Otherwise there are exactly two main N=-a and no other N. There are no
additional nonempty lines. The possibilities are:

| Main line costs | a forced by Gamma=a-1/2 |
|---|---|
| 4+4 | 1/2 |
| 4+5 | 1 |
| 4+6 | 1/2, both marginals0 |
| 5+5 | -1/2 or -1 |

For 5+5, the sums are a or a/2 per line; solving gives the stated negative
values. Consequently **positive a>1 requires f>=11**. For any b outside
the class prefix, B_ab<=2 when f(a)<=10: only one possible main half can
contain b, and its negative marginal is at worst -1. The Gamma0 cases
with extra balanced N are glue-free and give B=0. At f9 with nonzero
Gamma the class value is1, all glue is on its F5 line, and Gamma=1/2.

One nonzero source class plus a nonzero-marginal separating line costs
at least4+4+5=13. At f13 the together F4 lines have marginal0, so the
separating sum equals Gamma. A separating line with any nonzero glue
also costs at least5, even if its marginal is0; with two main lines this
still forces f>=13. These bounds permit balanced collisions on all lines.

#### 3.2 New tight two-class lemma

**Lemma S. Two nonzero classes at one attachment imply f>=13.**

If f<=12, the center-mass requirement of a nonzero actual star leaves at
most five N. Six N plus a star already cost14. Five N consist of a
three-N class and a two-N class. Different prefixes have just one shared
line, so their union has at least four N-lines, costing at least16.
For the same prefix but different groups, the four possible lines agree
and their V positions differ on every line (90 exact incidence pairs).
Five N have at least three lines; the tight distribution (2,2,1) costs
at least5+5+4=14, since two nonzero N at different V positions cannot
have F4.

Thus exactly four N remain, two per class, with coefficients (-a,-a)
and (-b,-b), and no other N. For different prefixes there are at least
three lines. To fit12 they must all be F4. The shared line must have
both N at the same V position with a+b=0; the other two glue-free stars
sum to a+b=0, contradicting center mass1.

For the same prefix, three lines already cost at least5+4+4=13, so both
classes must use exactly the same two lines. Their N occupy different
V positions. A shared F5 line has no stars and forces a=b; a shared F6
line has no stars and forces a=-b. A shared F7 line cannot occur when
a=b. These statements follow directly from S_i=l_i-r-nu_i, and are
independently checked by exact affine support masks. The only pairs
within cost12 are (5,5),(5,6),(5,7),(6,6). The middle pairs are inconsistent
and the remaining pairs have no actual star. Again center mass1 fails.

The bound is tight locally. The checker gives two value1/2 classes with
same prefix01 at center0, one F5 shared line and two F4 lines, f13 and
center mass1. It checks every V and h row at that center. This local
control is not a full PM2 identity.

#### 3.3 Inactive lightness and a dangerous sharing control

An inactive vertex with f<=9 has exactly one nonzero-marginal line, whose
sum is -1/2. Two such lines cost at least10; an extra nonempty zero line
costs at least6 in addition to5. Its B entries therefore belong to {0,1}.
All its actual stars have source equal to its half minus the center.
This statement concerns marginals, and does not assume nonnegative stars.

Naive main-star counting is unsafe. A local center0 control has main
class (M,01,0), D1437=1 and main N1400=N1419=-1. Glues
(453,454,457)=(-1/2,1/2,1/2) produce actual stars
(1157,1103,1132)=(1,1,-1), with f13. Add the balanced same-h-side pair
N1397=1, N460=-1. It removes stars1157 and1132, leaving just star1103=1,
and still has f13. Every V and h equation at center0 and center mass1
holds before and after. The balanced class has zeta0.

This explicitly refutes the claim that such a local unit-class/separator
pattern must retain two positive main stars. It is not a full identity or
a width27 counterexample. No argument below charges those removable stars.

### 4. Reduce all helper placements

If d>=3, the proved m2 charge Pi>=12+5d+z gives Pi>=27, and A>=1
excludes R27. If d=2, both vertices have two distinct nonzero D classes;
Lemma S gives Pi>=25+z. With z0 at most one D is in M, so A>=3 and
R>=28. With z1, C>=5; with z>=2, C>=2. All are larger. Thus **d<=1**.

If z1,d0, the other active vertex has zero zeta and f>=9. Thus Pi>=20,
A>=2 and C>=5. A>=3 already gives28. A2 forces the Z in M with value2;
its f>=11 makes Pi>=21 and again gives28.

If z1,d1, Pi>=20, A>=2 and C>=5. R27 requires Pi20,A2,C5,X0.
A2 forces both helpers into (M,K), beta2: an outside D or Z would add
an outside array; if K is not an M-edge, D is outside and no A2 placement
is possible. Private equality Pi20,h2 makes both active f8, all inactive
f5, hence both endpoint zetas1/2 and beta1, contradicting beta2.

Now z>=2. If a vertex has two nonzero classes, Lemma S gives
Pi>=ceil((13+8+20)/2)+h=21+h. Therefore h>=4 gives
R>=21+h+1+2>=28. If neither vertex has two nonzero classes, the only
possible h3 placement is D+Zu+Zv in one block; without D there are at
most two helpers. Thus **h<=3**.

For h3 and a multiple-class vertex, Pi>=24 and C>=2. Also A>=2.
For K in M this is immediate from the two untouched arrays. For K outside
M, A1 would require canceling its two crossing M-arrays with single Z of coefficient 2
at the two endpoints. The remaining helper lies outside those blocks and
creates a nonzero outside array; it cannot attach at the inactive partner
to share either crossing block. Hence R>=28.

For h3 with one nonzero class per endpoint, all helpers D+Zu+Zv occupy
one source-K block. Pi>=21 and A>=2 (at least3 if outside M), so C<=4.
The proved one-block lemma forces equal Z and zero flow. The exact
kernel (K) eliminates both Z at the cost of at most one new E, lowering
support. Matching lower27 excludes it.

Thus the entire Z branch reduces to **d0,z2**. The remaining Z=0 branch
is **one D**. This reduction explicitly covers same-source helper sharing;
it does not assert that every extra helper creates a different prefix.

### 5. Strengthen the one-block Z lemma to C<=5

If both endpoint Z lie in one block B, there is no other Z, and
C<=5, they are equal and every flow is zero. This hypothesis holds
both in the h3 single-block D+Zu+Zv use in section4 and the d0,z2
use in section6. For unequal Z, the projection right side is a nonzero
two-coordinate vector on U. Subtracting a possible multiple of c_B leaves
support2 or6, whereas one additional c has support4. At least two other
flow blocks are necessary. B needs two loops for unequal endpoint Z;
the two Z-free flow blocks cost two each. Hence C>=6.

For equal Z, projection is0. Distinct c columns are nonparallel (all990
pairs checked). Nonzero flow therefore needs at least three blocks. If
B flows, it needs at least two loops: its sole compatible E gives delta0,
and a sole L3 cannot satisfy both nonzero-Z S rows. Together with two
Z-free blocks this gives C>=6. If B does not flow, three Z-free blocks
give C>=6 in addition to B's loop. This proves the stronger lemma.

### 6. Two Z: same block or same attachment

For two Z in one block, Pi>=20,A>=2 gives C<=5. Section5 yields equal
Z=a and zero flow. Source balance (T) needs a separating endpoint, so
Pi>=23. If the block is outside M, beta=2a!=0 gives A>=4 and the budget
fails. If it is in M but beta!=2, A>=3,C>=2 also fails. The sole corner
is a=1,Pi23,A2,C2. Here Gamma=1/2 at both endpoints, the active cost sum
is at most22, so a separator13 and other endpoint9 are forced. The
separator sum is1/2, whereas (T) requires1. Contradiction.

If the two Z lie in distinct blocks but at the same attachment, that
vertex has two nonzero classes and the other active vertex has zero zeta:
f>=13,9. Thus Pi>=23. At most one Z-block is in M, so A>=3, and C>=2
gives R>=28. Hence distinct Z blocks must attach at different vertices.

### 7. Two Z: distinct blocks, different attachments

#### 7.1 Both outside M

Three M-arrays and two outside arrays give A>=5. The only budget is
Pi20,A5,C2,X0. Both active vertices are glue-free f8 with Z=1/2, every
inactive vertex f5, and each Z block has one L3 with |delta|=1/4.
All other blocks have beta0 and zero flow: d0 and the two Z-blocks
use the entire C2 budget. For an M-edge incident to an active vertex, H0<=1/2. Each helper block
contributes at most1/2 to beta-weight minus divergence: source1/2,
side1/4 plus at most1/4, absent0. Equation (F) would give w<=3/2<2.

#### 7.2 Both in M

Distinct M sources require K outside M. Their projection has nonzero
antisymmetric values on the two affected M-pairs, while every M c is
constant on each M-pair. A Z-free outside-M flow block is necessary,
so C>=4. If A2, one Z is2, forcing Pi>=22; if A1, both are2, forcing
Pi>=23. Both exceed27. The only corner is Pi20,A3,C4,X0: both Z=1/2,
inactive f5, one L3 in each Z block and two loops in a single outside
flow block B'. Let f be the unaffected M-edge. Projection requires c_B'
constant on f, hence its group contains f and neither affected M-edge.
This is checked on all270 applicable group/block incidences. Thus B' has
no edge divergence at an affected source e. There H0<=1/2; helper terms
are 1/2+1/4; the other M-block contributes divergence of magnitude1/4.
Again (F) gives w<=3/2<2.

#### 7.3 One in M, one outside

Let the M Z=a attach at u, source e={u,p}; the outside Z=b attaches
at v. A>=3. For A3 the M array must cancel, so a=2 and Pi>=22. R27
forces Pi22,C2,X0, with f(u)<=12, f(v)<=9, inactive f5. For A4,
Pi<=21,C<=3: both active f<=10, inactive f<=7. A5 forces
Pi20,C2,X0 and satisfies these same constraints. A>=6 exceeds R27.
This includes Pi20,A4,C2,X1 and the extra-loop C3 corner.

No Z-free nonzero-flow block can fit C<=3, since the two Z blocks already
cost two loops. Since f(u)<13, u has no nonzero-marginal separator for e,
and B_up=1-2a. Classify e in the outside group:

* If e is its source, p=v and both endpoints have no separator. Then
  H0_e=1-a-b, helper=a+b and divergence0, giving w_e=1 instead of2.
* If e is a side, both possible c columns are constant on e. Projection
  there has the nonzero antisymmetric contribution of the M Z and no
  outside-Z source contribution. Contradiction.
* If e is absent, divergence0 and helper=a. Equation (F) requires
  B_pu=3. An inactive p has B<=1 by f<=7. If p=v, u lies outside its
  prefix and the collision-safe f<=10 bound in section3.1 gives B_vu<=2.
  Neither can reach3.

This closes every two-Z placement, including signed values, idle loops
and the possible mixed column. Thus **Z cannot occur at R27,m2**.

### 8. Single D: zero flow and the private22 obstruction

Now d1,z0 with D=a!=0 and source K. Nonzero flow costs C>=6.
If its group is outside M, A>=4 and Pi>=19. If it is in M and a!=1,
A>=3 and Pi>=19. If a=1 in M, both Gamma1/2 force Pi>=20 and A>=2.
All give R>=28. Hence flow is zero, and (T) applies.

A separator costs13 and the other endpoint8, so Pi>=22. **Pi22 is
impossible.** If a!=1/2, both Gamma are nonzero. The total f budget42
forces separator13, other9 and inactive5. The other9 forces a=1, while
the separator sum Gamma=1/2 cannot equal NT=+/-1.

If a=1/2, Gamma0. A separator13 would have marginal0, so separator>=14;
the other endpoint8 and all inactive5 are forced. Three main N-lines
or an extra together line cost at least12 before a separator5 and cannot
fit14. Thus there are precisely two main N=-1/2 on their main lines,
with together cost at most9, allowing any balanced collisions. Their
total marginal is0,1/4 or1/2 by F4/F5, and cannot equal -NT=+/-1.
Therefore **Pi>=23** for every single D.

### 9. Single D outside group M

A>=4 forces precisely Pi23,A4,C=X0. A separator13 and other8 imply
each inactive f<=8; hence every inactive B row is in {0,1}, by section3.3.
If K is an M-edge, the two other M-edges are absent from the D group,
and (F) asks H0=2 despite H0<=1. Contradiction.

If K is not an M-edge, let f be its unaffected inactive M-edge. Unless
f is a side of the D group, (F) again asks H0=2. If it is a side,
H0_f=2-a<=1, forcing a>=1. The other side of the D group pairs the two
inactive matching partners of u and v; call it j. It is not an M-edge.
Both endpoints of j are also light, so H0_j>=0. But (F) at j gives
H0_j=-a, forcing a<=0. Contradiction. This short signed inactive-edge
obstruction is also present in Part D; here its lightness premise has
first been derived from Pi23, including any balanced N at u,v.

### 10. Single D in M, a!=1

A>=3 and Pi>=23 exclude C>=2; C1 is impossible. For X0 all alternatings
are in group M: the three nonzero M arrays already cost3, and any
outside zero-total array would add2, while A<=4. The full403-column
dual (D21) excludes every such case.

The only X1 corner is Pi23,A3,C0. The separator charge gives active
total at least22:13+9 for Gamma!=0, or14+8 for Gamma0. Thus inactive
f<=7 and H0 at either inactive M-edge is in {0,1/2,1}. Equation (F)
again forces a in {1,3/2,2}; the nonunit cases are3/2 or2.

Positive a>1 makes the other endpoint cost at least11. The total44
forces separator13, other11 and all inactive5. The source balance NT=1
then forces Gamma=1 and a3/2. The separating F5 line has nonzero
glues of magnitude1. Such a glue cannot cross the other active vertex:
that vertex has two main N-lines of cost at least8, and a nonzero
separating glue would add a line of cost at least5, exceeding f11.
It cannot cross an inactive f5 vertex either, whose glues have magnitude
1/2. Contradiction. This proof allows a balanced N on the separator;
it charges the separating glue and main lines jointly, not their stars.

### 11. Single D in M, a=1: private23,24,25

#### Pi23

Both Gamma=1/2 and the active total is at most24. At most one endpoint
has a nonzero-sum separator. Its cost cannot be13, since that would give
separator sum1/2 rather than NT1. Thus it costs14 or15 and the other
endpoint at least9. Three main lines or an additional together line would
cost at least12 before a separator5, exceeding15. Thus there are exactly
two main N=-1, no additional together line, and together cost<=10.
All F4/F5/F6 main
marginals are nonnegative, including collisions. Their required sum is
Gamma-NT=-1/2. Contradiction independent of A,C,X.

#### Pi24

A>=2,C1 impossible force C0 and A<=3. All alternatings are in group M:
two nonzero M arrays already need2, and any outside zero-total array
would require2 more. X1 is allowed here.

The total f budget is46. The separator13 and other nonzero-Gamma9
give active total>=22, hence every inactive f<=9. This establishes
lightness **before** using inactive-edge equations. Their B entries are
0 or1. At each inactive M-edge, (F) gives H0=2-a=1, so both directed
entries are1. Every inactive center's unique nonzero-marginal half
therefore contains its inactive matching mate.

No actual star can belong to block (M,K). An active center is itself
in K and cannot have source K. An inactive center's star source is its
half minus the center, which contains an inactive matching mate, and
therefore is not K. This conclusion uses literal sources and holds
even for signed stars and the balanced-N sharing control of section3.3.
All group-M alternatings are disjoint from the four block-alternating
columns of (M,K), as checked on all45 group/block incidences.
Thus both sums on the left of (B) are zero, but beta=2. Contradiction.

#### Pi25 or higher

Pi25 forces A2,C=X0 and every alternating in group M. The full dual
(D21) gives -1=0. Pi>=26 plus A>=2 already exceeds R27.

This completes the single-D branch and therefore the m2 theorem.

## I. Three attachment vertices



Use the proved general identities and costs, and the helper-array
rigidity Omega=-2e_M in Part D. Set h=d+z. Then

    R=Pi+A+C+X=27,   Pi>=21,   Pi>=20+h,
    2Pi=sum f+2h.

The three inactive vertices cost at least 15 in sum f. Every active
vertex costs at least 8, or 9 when Gamma!=0. A vertex with zero total
helper classes has Gamma=-1/2. A surviving D requires both source
endpoints in K. Thus K either contains one M-edge plus one endpoint of
another, or it meets all three M-edges once. At most one M-edge lies
wholly in K.

### 1. Three elementary facts used in the cover

**A value-2 class costs at least ten at its endpoint.** If it is the only
nonzero helper class, the proved value-2 lemma gives f>=11, including
balanced N classes. If another nonzero helper class is present, there
are at least four N and at least one actual star, giving f>=10. We use
this weaker uniform bound when extra helpers may be present.

**Two different-source classes give f>=12, or f>=13 if Gamma!=0.**
Each needs two N on distinct private lines; their line sets share only
one line. Thus at least three nonempty N-lines occur, each costing at
least four: two colliding N already cost4. Equality12 makes all of
them F4 and glue-free, and leaves
no other nonempty line. Thus a nonzero Gamma requires at least 13.
This permits arbitrary balanced classes and collisions; adding them
cannot erase the four nonzero main-class coefficients.

**Matching-group flow vectors are constant on each M-edge.** A block
with group M has c_B=1_W-1_V, with U,V,W the three edges of M. Its vector
is constant on the two coordinates of each M-edge. If all Z and all
nonzero flow blocks lie in M, the projected right side is a sum of
antisymmetric endpoint differences, separately on the three M-edges.
Equality with the left side therefore forces the endpoint Z values in
each M-block to be equal. In particular, a single Z on an M-edge is
impossible in such a configuration. D is absent from this projection.

### 2. No Z: all D cases

At least two M-arrays are untouched because at most one M-edge is inside
K. Hence A>=2. Nonzero flow would need C>=6, giving R>=21+2+6=29.
The flow therefore vanishes. We do not infer the absence of idle loops
until a support budget does so.

#### One D

Let its coefficient be a and its source U. Source balance gives
NT_U=w_U-1=+1 or -1, including all balanced N at the third active vertex.
(T) involves only the line sums at the source endpoints. Balanced N
elsewhere, including at the third active vertex, do not enter it. A
nonzero-sum separator costs at least5 even with balanced N. At f13 its
sum equals Gamma because the two cost4 main lines are glue-free.
Some D endpoint has a nonzero separating marginal and f>=13, the other
has f>=8, the third active vertex f>=9, and inactive sum f>=15.
Thus Pi>=24. A D outside M would leave all three M-arrays and add its
own nonzero array, giving A>=4 and R>=28. Therefore its block is (M,U).

At Pi24, the D endpoint costs sum to at most 22. If a!=1/2, both endpoint
Gamma values are nonzero, so the tight endpoint costs are (13,9), with
the third active f9 and every inactive f5. The proved f9 marginal law
forces a=1. The f13 endpoint's separating marginal is then Gamma=1/2,
and the other endpoint has no separator. This contradicts NT_U=1.

If a=1/2, an f13 endpoint cannot have a nonzero separator because its
together F4 lines are glue-free and Gamma=0. Thus the tight endpoint
costs are (14,8). The f14 endpoint alone must have separating marginal
1 and together marginal -1. Its two main N=-1/2 together lines cost
at most nine in total, after allowing at least five for the separator.
There cannot be a third main N line or another nonempty together line.
The full collision-safe F4/F5 lemma gives total together marginal in
{0,1/4,1/2}, a contradiction. This is precisely the stated endpoint
argument (proved the prior component R3), with the third active vertex's
balanced classes left allowed.

At Pi25, the budget forces A2,C=X0 and a=1. The endpoint cost sum is at
most 24. If a separating endpoint has f13, it supplies only 1/2, so the
other endpoint would also need a separator, costing at least 26 together.
Otherwise the unique separating endpoint has f14 or f15 and supplies
separating marginal 1. Its together marginal is -1/2. The two main
N=-1 lines cost at most ten; extra together lines cannot fit. The
collision-safe F4/F5/F6 possibilities give only nonnegative total
together marginal, contradiction. This is the proved collision-safe
R2b main-line argument, including balanced N. Larger Pi cannot fit A>=2.

#### Two D with the same source

Each endpoint has two nonzero classes and at least four N, hence f>=10.
Source balance still forces a nonzero separator at one endpoint. Its
classes use at least two together N-lines costing at least eight, and
the separator costs at least five, so that endpoint has f>=13. The third
active vertex has f>=9. Thus

    Pi>=ceil((13+10+9+15)/2)+2=26.

The distinct D blocks cannot both be in M; their array count is at least
three. This already exceeds R27.

#### Two D with different sources

The sources form a path on K. At the shared endpoint there are two
different-source classes, so f>=12 by section 1. Each leaf costs at
least eight; together with inactive sum 15 this gives Pi>=24. If neither
D is in M, A>=5. If one is in M without canceling its array, A>=4.
Therefore a possible R27 identity needs Pi24,A3,C=X0 and D=1 in an
M-block; call the other coefficient b.

The M leaf has Gamma=1/2 and f>=9. The shared vertex has Gamma=b+1/2,
while the other leaf has Gamma=b-1/2. These two Gamma values cannot both
vanish. If the shared Gamma is nonzero its cost is at least 13; otherwise
the other leaf costs at least nine. In either event the three active
costs sum to at least 30: 13+9+8 or 12+9+9. Consequently

    Pi>=ceil((30+15)/2)+2=25,

contradicting Pi24. This argument uses no prescribed coefficient b and
does not assume that a shared f12 line configuration is feasible.

#### At least three D

The proved D-degree bound restated in Part B gives Pi>=26. At least two M-arrays still
need A>=2, so R>=28. This completes Z=0.

### 3. Z present: preliminary restrictions and all A<=1 cases

For z=1, Pi>=22. When d>=1 this follows from Pi>=20+d+z. When d0,
the two helper-free active vertices cost at least nine each, the Z
endpoint at least eight, and inactive cost is 15, giving Pi>=22 again.
At least one M-array is unchanged: if K contains an M-edge, D can affect
only that one and the Z at most one further array; if K meets all three
once, D affects none and Z at most one. Thus A>=1. The *proved* bound
C>=5 for a single Z already yields R>=22+1+5=28. This note does not
need the stronger C>=6 lemma from Part G.

Hence z>=2. Support-minimality against the proved global lower30 and
the literal S rows give C>=2. Also Pi>=20+h>=22 and Pi<=25. We first
rule out A0 and A1, to avoid missing high-helper boundary cases.

**A0.** K must meet every M-edge once, since an untouched edge has a
nonzero array. All three arrays must cancel. No M-edge is contained in
K, so each cancellation requires its single endpoint Z=2. There are
at least three helpers and each active vertex has a value-2 class.
Section 1 gives f>=10 at all three vertices, hence

    Pi>=ceil((15+30)/2)+3=26,

already impossible with C>=2.

Thus A>=1 and Pi<=24. If A1, at least two M-arrays cancel.

**A1 and K meets all three M-edges once.** At least two distinct vertices
carry Z=2. If h2, they each cost at least eleven and the third active
vertex costs nine, giving Pi>=25. If h>=3, use costs 10,10,8 to obtain
Pi>=ceil((15+28)/2)+3=25. Therefore no Pi<=24 case fits.

**A1 and K=U union {w}, U in M.** Let V be w's M-edge and W the fully
inactive M-edge. The W-array consumes the sole alternating column.
Consequently both U and V must cancel, and Z_w=2 in (M,V).

If h2, z>=2 means the other helper is a Z=2 at an endpoint of U. The
costs 11,11,9 again give Pi>=25. If h>=4, the value-2 endpoint costs
at least ten and the other active vertices at least eight, giving
Pi>=ceil((15+26)/2)+4=25. It remains h3.

All helpers then lie in the two canceled M-blocks. A nonzero outside
beta would require an additional alternating; a zero-beta outside block
would need at least two helpers besides the two cancellations. The
V-block contains only its Z_w=2. The U-block contains the other two
helpers, either D plus Z or the endpoint Z pair. The w cost is at least
eleven and the other two active costs at least eight, so Pi>=24. The
only budget is Pi24,A1,C2,X0. There are Z in both M-blocks, so both
loops are confined to them. Their flow vectors are constant on M-edges,
but the single Z at w gives a nonzero antisymmetric projection on V.
Section 1 gives a contradiction. All A1 cases are now excluded.

### 4. Z present: the remaining Pi22 and Pi23 budgets

We have A>=2,C>=2 and Pi>=22. Thus only

    Pi22 with A+C+X=5, or Pi23,A2,C2,X0

can remain. The following cases exhaust h=d+z and helper placements.

#### Pi22

The bound Pi>=20+h forces h2, so d0,z2. Both Z attach at different
vertices: attachment at one would cost at least ten there and nine at
both other active vertices, exceeding the active budget 25. The exact
costs are f8 at the two Z endpoints, f9 at the third active vertex,
and f5 at every inactive vertex. Thus both Z equal 1/2.

No M-array cancels, so A>=3, forcing A3,C2,X0. An outside-M Z block
would add a nonzero array, so all Z blocks are in M. If they are distinct,
the two loops are in those blocks and the projection contradiction of
section 1 applies. If both Z are in one block, the proved one-block
lemma gives zero flow. Source balance then contradicts the glue-free
f8 endpoints: NT_U=0 whereas w_U-1=1. This closes Pi22.

#### Pi23, h2

Again d0,z2. At least one M-array cancels. If the Z are in different
blocks, an outside-M Z would add an array besides the two remaining
M-arrays, exceeding A2. Thus both are in distinct M-blocks, excluded by
the same projected-flow argument with C2.

If the Z are in one block, the one-block lemma with C2 forces equal
nonzero Z and zero flow. Its beta is nonzero, so A2 then forces that
block to lie in M and beta to equal two. Both Z are therefore one.
Source balance requires a separator, giving active costs at least
13,9,9 and inactive cost 15. Including the two Z gives Pi>=25, not 23.

#### Pi23, h3, d0,z3

At most two Z blocks are possible because C2. If all Z are in M,
one block has a lone endpoint Z, and projection gives a contradiction.

If an outside block has beta zero, it contains two opposite endpoint
Z. Such a block requires at least two loops, besides the loop for the
remaining Z; C>=3, impossible. If an outside block has beta nonzero, C2 permits at most two Z blocks.
There is at most one other Z block, which can cancel at most one M-array.
Two M-arrays and the outside array therefore require A>=3, contradiction.

#### Pi23, h3, d1,z2

If both Z lie in one block, the one-block lemma makes them equal and
the flow zero. If that block and the D are distinct, their arrays need
at least three alternatings: at most one M-array can cancel and a
nonzero outside array remains. Thus A2 forces all three helpers into
the same M-block. The equal-Z kernel move deletes both Z and creates
at most one E, while modifying an existing D, strictly lowering support.
That contradicts the proved global width-30 minimum.

Now put the two Z in distinct blocks. If both are in M, their two loops
are confined to M and projection excludes them. If both are outside M,
at most one M-array can cancel via D; the two outside Z arrays cannot
both have beta zero using a single D. This exceeds A2. Hence exactly
one Z is in M and the other outside.

If D lies in the same M-block as the M Z, the other two M-arrays plus
the outside Z array need A>=3. If D lies in a different M-block, the
only way to fit A2 is D=1 and M Z=2, canceling two M-arrays. The
outside Z gives an additional class at some vertex. The value-2
endpoint costs at least ten and the other two active vertices at
least eight, giving Pi>=ceil((15+26)/2)+3=24, contradiction.

If D is outside M, the M Z must be 2 to leave exactly two M-arrays.
The outside beta must vanish, so D and the outside Z share a block
with Z=-2D. Its two endpoint classes are nonzero and opposite.
If its source contains the M-Z endpoint, active costs are at least
10,8,9; otherwise they are at least 11,8,8. Either way Pi>=24.
Every d1,z2 placement is excluded. Finally h>=4 would force Pi>=24,
so it cannot occur in Pi23. This completes all Z cases.
