## 1. Counts and an allowed local normalization

All retained coefficients are scaled by three. Every vertex has centered
star mass one. A matching correction has unscaled coefficient 2/3 on each
of its three edges, so its scaled ordered slots have target -2.

Use the following support counts:

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

Here N is nonalternating P4 support, Z zero-loop support, D doubled support,
A alternating support, E double-loop support, and X mixed support. R is
retained support; total width in the matching branch is R+3. Let m count
vertices carrying at least one nonalternating direction as its attachment
vertex. The proof assumes R<=23 and derives a contradiction.

For a loop block B with source U={u,v} and other pairs V,W, write

    zeta_B(u)=Z_B(u)+D_B,       zeta_B(v)=Z_B(v)+D_B,
    z_B=zeta_B(u)+zeta_B(v).

The h equations say that each of the two nonalternating side sums at
(B,u) is -zeta_B(u), and similarly at v. Thus nonzero zeta at an endpoint
requires nonalternating directions attached there. Each such class needs
at least two selected N directions, one on each side; classes are disjoint.

Use the full-matrix relation in the canonical quotient by the free
feature-linear term ell.m:

    D_B - Z_B(u) - Z_B(v) + E_B = 0

This relation also has a direct functional explanation with consistent
raw orientations. Set K=m_V+m_W, D0=2m_U-K, E0=x_u+x_v-K, and
Z0(u)=m_U+x_u-K, Z0(v)=m_U+x_v-K. Pointwise the unordered multisets
{Z0(u),Z0(v)} and {D0,E0} agree: on x_u>=x_v the first Z is D0 and
the second is E0, and on x_v>=x_u they swap. Their hinge combination is
therefore zero. The stored quadruples have a common orientation in each
block (15 positive and 30 negative); all 45 full-matrix relations are
rechecked. An independent projective sign reversal adds only a
feature-linear term, which is absorbed into ell.m. This corrects a sign
error in the optional functional explanation in the the prior component reply;
the exact kernel relation and support move are unchanged.

This permits removing D_B whenever at least one endpoint zeta is zero. If, for example,
zeta_B(v)=0, the coefficients have form

    (D,Z_u,Z_v,E)=(d,alpha-d,-d,e).

Replace them by (0,alpha,0,e-d). If d!=0 and alpha!=0 the old support is
at least two and the new support at most two. If alpha=0, the old support
is at least three and the new at most one. If d=0 nothing changes. This
normalization preserves zeta, stars, glues, N, A, corrections and m, and
never increases **total** support, even if it introduces E.

After normalization:

* a nonzero D has two nonzero endpoint zetas, hence two N-active endpoints;
* Z vanishes at N-inactive vertices;
* when m=1, D=0;
* when m=2, every D has the same source pair consisting of those vertices.

This is narrower than deleting D without qualification, which is not
support-nonincreasing in general.

## 2. Private vertex charges, including their equality case

A private line l has three glue entries, sum s, and three V equations

    S_y=l_y-s-nu_y,

where nu_y is the sum of its two possible N coefficients. Define

    F(l)=support(l)+2*stars(l)+2*N(l),
    pi(u)=1/2 sum_(lines of u) F(l) + Z(u) + D(u)/2.

Each glue occurs on two lines; each star and N on one; each D has two
source endpoints. Consequently sum_u pi(u)=Pi exactly. We use the
previously checked facts:

* a nonzero-sum line has F>=5;
* a line containing N has F>=4;
* the two N directions on one h side have different private lines, and
  the four directions in both sides of a class occupy four different lines;
* with Gamma(u) the sum of its glue marginals and
  Psi(u)=sum_B zeta_B(u), one has Gamma(u)-Psi(u)=-1/2.

Every N-active vertex has selected N on at least two private lines. For
nonzero zeta, the two opposite h sides force distinct lines. When all
zetas vanish, a selected N belongs to a zero-sum h side and must have
its cancelling partner on a different line. Thus the conclusion also
holds for a cancelling class. If
some zeta at u is nonzero, its line charge is at least four and its Z/D
charge at least 1/2, giving pi(u)>=9/2. If all its zetas vanish, normalized
Z/D vanish there and Gamma(u)=-1/2. The two N lines cost at least four
each, and a nonzero-sum line raises the total line cost to at least nine.
Again pi(u)>=9/2. An N-inactive vertex has pi(u)>=5/2. Therefore

    Pi >= 15+2m.                                            (1)

For matching corrections there is a useful equality consequence:

    Pi=15+2m  implies  A>=3.                                (2)

Indeed, equality fixes every vertex charge at its indicated minimum. At
an active vertex with nonzero zeta, its helper charge is exactly 1/2:
there is one D and no Z, and its two N lines each have F=4. A line with N
and F=4 has no glue entries. Indeed, the only glue-bearing integer-count
candidate is (glues,stars,N)=(2,0,1). For l=(p,q,0) with p,q nonzero,
l-s*1=(-q,-p,-p-q) has at least two nonzero entries; one N coefficient
cannot cancel all of them. Thus Gamma(u)=0 and Psi(u)=1/2,
so its unique D coefficient is 1/2. Every such D joins two vertices of
this kind, and their source pairs form a matching because no vertex can
carry a second D. There are no Z directions. Accordingly each aggregated
source zeta sum is zero or one. By the slot law below, every edge of the
correction matching still has a nonzero alternating isolated sum, either
-2 or -1. At least three alternating directions are required.

## 3. Exact group, slot, and projected-flow identities

For a perfect matching G and an edge U of G, let t_U^G be the total of its
four alternating directions with isolated pair U. In the block (G,U) the
paired T/V/h equations give a common value

    Omega_G = t_U^G-z_(G,U),       independent of U in G.     (3)

The checker supplies exact combinations of at most 18 local equations
for the two equalities in each of all 15 groups. In particular, writing
I_e for the total of all alternating directions isolated at e,

    I_e = -2*indicator(M)_e + sum_(B:source=e) z_B,
    sum_(G contains e) Omega_G = -2*indicator(M)_e.           (4)

Each alternating direction has exactly one isolated pair, so
A>=support(I). Since the right side in the second equation is nonzero,
at least one Omega_G is nonzero.

Let L denote L3 support alone. Equations (3) and the S rows imply

    2A+L >= 6-m.                                            (5)

To see this, fix a group with Omega_G!=0. For each of its three edges U,
if t_U^G!=0, its array needs an alternating column, which can account for
at most two N-inactive endpoints. If t_U^G=0, then z_(G,U)!=0, so at least
one endpoint is N-active. If the other is inactive, its zeta is zero and
the difference of the two S rows forces an L3 direction in that block:
the Z coefficients differ, while D and E cancel in the difference. The
three blocks and the three alternating arrays are disjoint. Counting the
6-m inactive endpoints proves (5).

There is also a coordinate projection of the general flow law.
Orient a block's flow edge V->W and let its projected column be
c_B=1_W-1_V. For matching corrections,

    2 sum_B delta_B c_B
      = sum_(B:source={u,v}) (Z_B(v)-Z_B(u))*(e_u-e_v).      (6)

For an explicit derivation, put beta_B=z_B, Sbeta=sum_B beta_B,
Ubeta(w)=sum_(B: w in source(B)) beta_B, and
Psi(w)=sum_B zeta_B(w). Literal V substitution and h elimination give

    deg(H)_w  = 2 + Sbeta/2 - Psi(w) - Ubeta(w),
    deg(nu)_w = Sbeta/2 - 3*Ubeta(w)/2.

Subtracting the matching degree 2 gives Ubeta(w)/2-Psi(w), which is
half the sum of partner zeta minus own zeta over source pairs at w.
The D terms cancel. This proves (6), with the factor two as written.
The checker independently derives all six coordinate identities. Each c_B
vanishes at its two source vertices. If Z=0, the projected flow is zero,
and the 45 distinct nonparallel c_B imply b=0 or b>=3, where b counts
nonzero delta_B. With Z=0 every nonzero flow block needs at least two
L3/E directions, so C>=2b. In general only a block without Z enjoys that
additive two-column charge; Z has already been counted in Pi.

A block with Z present needs at least one L3/E direction by its S rows.
If exactly one endpoint Z is nonzero, it needs an L3 direction. If there
is exactly one Z direction in the entire identity, then

    C>=3.                                                   (7)

One loop term is needed in its block. Equation (6) has a nonzero entry
at that Z's source vertex, whereas that block's c_B is zero there. Some
other flow block is therefore nonzero; that other block has no Z and
needs at least two L3/E directions. These costs are disjoint.

## 4. All zeta-zero configurations and m>=3 are excluded

When N=0, h forces every endpoint zeta to vanish, and the allowed
normalization removes D and Z. Then Pi=P, effective and actual stars
coincide, P>=15 and A>=3. If H=2M, Appendix A gives P>=24. Otherwise
flow divergence is nonzero. Its projection is zero by (6), so b>=3 and
C>=6, giving R>=15+3+6=24. This closes N=0 directly, without the
stronger width-30 or two-partition theorems.

For the rest of this section assume N>0.
First suppose every zeta vanishes. Normalization removes all D and Z.
Define effective stars by the private formulas S0=l-s. The actual stars
are S=S0-nu, so their effective private support P0 is at most stars+glues+N,
which is Pi<=23. Each N has a unique V row; an effective-star position
absent from the actual support requires an N in that row, so different
new positions require different N directions. The h equations give
N mass zero at each center, so S0
has center mass one. They also show H(actual stars)=H(S0): the two N
directions on each h side have the same pendant pair and opposite total.
The nonalternating flow source vanishes. By Appendix A,
H(S0)=2M would require P0>=24. Hence flow divergence is nonzero; b>=3,
C>=6, and (4) gives A>=3. Since N>0 implies Pi>=17,
R>=17+6+3=26, contradicting R<=23.

Now suppose m>=3. If m>=5, (1) already gives Pi>=25. If m=4, Pi>=23;
equality forces A>=3 by (2), while Pi>=24 is already sufficient. If m=3,
Pi>=21. At Pi=21, (2) gives A>=3. At Pi>=22, (5) forces at least two
directions among A and L, since 2A+L>=3. In every case R>=24.

It remains to treat m=1 or 2 with some nonzero zeta.

## 5. One attachment vertex

Let u be the only N-active vertex. Normalization gives D=0, and all Z
directions are at u. Write z for their number; z>=1. Two matching edges
avoid u, so (4) gives A>=2. Every Z lies in a distinct block requiring a
loop term, so C>=z. Its distinct h classes require N>=2z at u. There is
also at least one star at u. The other five vertices contribute at least
25/2 to Pi. Thus

    Pi >= ceil(25/2+1+2z+z)=14+3z.

For z>=2, R>=14+3z+2+z>=24. Consider z=1. The line charge at u is at least
four, giving Pi>=ceil(25/2+4+1)=18. Equation (7) gives C>=3. If A>=3 the
proof is complete. If A=2, its two isolated sums must cover the two
unaffected matching edges, at coefficient -2 each. Equation (4) forces
the sole Z source to be the remaining matching edge, with Z coefficient 2.

Only Pi=18 could then leave R<=23. Gamma(u)=3/2 already forces its line
cost to be at least nine, because cost eight has no glue. The total line
cost is 2(Pi-Z)=34; the other five vertices use at least 25, exhausting
the budget. Center mass one also requires a star at u. Thus equality gives
total private line cost F=9 at u and cost-five lines at the other vertices.
Both h sides of its nonzero class have sum -2. With F<=9 there are exactly
two N directions, one on each of two lines: three from the class would
occupy three lines of cost at least four, and extra zero-zeta classes
come in pairs, giving at least four N plus a star and hence F>=10.

The two lines have costs five and four. The cost-four line has no glues.
For a line with one N coefficient n and F=5 there are just two possibilities:

| Glues | Stars | Glue-line sum |
|---:|---:|---:|
| 1 | 1 | -n |
| 3 | 0 | -n/2 |

This follows directly from F=g+2s+2=5 and S=l-(sum l)*1-n*e_y. With
n=-2, Gamma(u) is therefore 2 or 1. But the center identity requires
Gamma(u)=Psi(u)-1/2=2-1/2=3/2. This contradiction closes m=1.

## 6. Two attachment vertices: all but one helper case

Let U={u,v} be the two N-active vertices. All D directions have source U.
Write d for their count, z for Z count, and h=d+z>=1. The inactive vertices
contribute at least ten to Pi and the two active line charges at least
eight, so

    Pi>=18+h.                                               (8)

At least one matching edge avoids both active vertices, hence A>=1.
When all helper sources are U, only the isolated sum at U can change, so
at least two matching edges are unaffected and (4) gives A>=2.

If d>=2, normalization ensures two nonzero endpoint zetas for every D.
Thus N>=2d at each active vertex, in distinct classes. Each vertex also
needs at least one star. Their line charges total at least 2+4d. Hence
Pi>=12+4d+h. For h=2 this gives Pi>=22 and A>=2; for h>=3 it gives
Pi>=23 and A>=1. These cases have R>=24.

We may assume d<=1. The remaining counts are:

| Helpers | Exclusion |
|---|---|
| h>=6 | Pi>=24. |
| h=5 | Pi>=23 and A>=1. |
| h=4 | Pi>=22, A>=1, and some Z forces C>=1. |
| h=3, d=0 | Three Z require at least two blocks: Pi>=21, A>=1, C>=2. |
| h=3, d=1 | If the two Z are in different blocks, use C>=2. If in one block, its source is U and all helper sources are U, so A>=2 and C>=1. |

Two Z directions in one block occupy its two source endpoints. Both
must be N-active, so that source is U. Each D contributes one unit of
support to Pi, split as one half-charge at each endpoint. A D source U
and any Z source intersect in an active vertex, so they can affect at
most one edge of a perfect matching. These observations justify the
source and charge assertions in this case table and the next cases.

For h=2 with d=1,z=1, Pi=20 would require both active line charges to be
exactly four. They have no glue entries, so Gamma(u)=Gamma(v)=0 and
Psi(u)=Psi(v)=1/2. The D contributes equally at both vertices, but the
single nonzero Z adds at only one: impossible. Thus Pi>=21, A>=2 (one D
source U plus one Z source can affect at most one matching edge), and
C>=1, giving R>=24.

For h=2 with d=0,z=2, Pi=20 likewise forces Gamma(u)=Gamma(v)=0. There
must be one Z at each active vertex, each of coefficient 1/2. No matching
isolated sum can then cancel its target -2, so A>=3 and C>=1. If Pi>=21,
two Z blocks give C>=2 and A>=1; a single Z block has source U and gives
C>=1 and A>=2. Again R>=24.

For h=1 consisting of one Z, the active vertex without zeta has line
charge at least 9/2; the other has at least four plus its Z charge one.
Together with the inactive charge ten, Pi>=ceil(39/2)=20. Here A>=2 and
(7) gives C>=3. This case is also excluded.

Only one D and no Z remain.

## 7. One doubled coefficient: the tight private cases

Let its coefficient be a!=0 and its source U={u,v}. At both active
vertices Psi=a and Gamma=a-1/2. Their only nonzero-zeta class has the two
h side sums -a. We have Pi>=19 and A>=2. Because Z=0, any nonzero flow
needs at least six loop directions, contradicting R<=23. Thus b=0.

We use the following small-cost facts at an active vertex. Let f be the
sum of its private line costs, excluding its D half-charge.

For a line without N, the possible nonzero-sum costs are 5,7,8,9;
zero-sum costs are 0,6,9. In particular there is no stray pure line of
cost one or two. Together with the distinct h/V class ownership this
justifies the small-cost classifications below.

* f>=8. If f=8, all its glue entries vanish.
* If f<=9, there are exactly two N directions of coefficient -a, on the
  two sides of the nonzero class. The reason is the same as in section 5.
* If f=9, the cost-five/cost-four classification gives
  Gamma in {a,a/2}. Thus Gamma=a-1/2 is possible only for a=1, with a
  three-entry glue line of pattern (1/2,1/2,-1/2).
* If a=1/2, Gamma=0. A glue-bearing vertex has f>=10. At f=10 there are
  still exactly two N directions: four N plus a star and a glue would
  cost at least eleven, and three from the nonzero class use three lines.
  The only possible glue-bearing form is one cost-six line and one
  cost-four line. The former has glue entries +1/2,-1/2,0, sum zero; its
  N coefficient -1/2 cancels the negative entry. Two cost-five lines
  would instead give positive total Gamma.

The literal V incidence has one further consequence: the private half
of every N direction in this class contains both u and v. For a path
p-a-b-w with isolated pair Y, its V row has center a and private half
{p,a,b}. The class prefix pair {p,a} is therefore in one half. The
incidence is also checked for all 360 N columns from their actual V rows
and path metadata; no additional support classification is used.

**Case a=1/2.** Equation (4) gives A>=3, so Pi<=20. If both active vertices
have no glues, H_G(U)=(B_uv+B_vu)/2=0. The exact H substitution and nonalternating source
then give (H-nu)_U=1. This cannot equal 2*indicator(M)_U, contradicting b=0.

If an active vertex has glues, its f is at least ten. The budget Pi<=20
has only two units of line-cost excess over the baseline f=8,8 and four
inactive costs five. Thus one active vertex has precisely the last
glue-bearing form above, the other has no glues, and all inactive lines
have their cost-five equality pattern with sum -1/2. The glue-bearing
line lies in a private half containing u,v. Its positive entry +1/2
crosses an inactive line of sum -1/2 and cost five. That line must have
pattern (+1/2,-1/2,-1/2), so all three crossing entries are nonzero,
including the entry on the other active vertex's line. This contradicts
that vertex having no glues.

**Case a!=1/2.** Both active Gamma values are nonzero, so f>=9 at both and
Pi>=20. Pi=20 would force f=9 at both and equality on all inactive lines.
The preceding classification then forces a=1 and a full glue line of
pattern (+1/2,+1/2,-1/2) at each active vertex. Its two positive entries
cross inactive cost-five lines of sum -1/2, forcing those lines to have
three nonzero entries. Since u and v belong to the same private half,
both active glue lines and the third inactive line belong to that same
private block. Its half containing u,v has marginal sum
1/2+1/2-1/2=1/2, while the opposite half has three inactive marginals
summing to -3/2. A matrix cannot have different total row and column sums.
Thus Pi>=21.

Unless a=1 and U belongs to M, (4) requires A>=3, already giving R>=24.
The only remaining possibility is

    a=1, U in M, Pi=21, A=2, C=X=0.                         (9)

## 8. A 21-row exact contradiction for the final case

Let G0 be the matching group of the sole D block. Every other group has
z_B=0. With only two alternating directions globally, (3) forces
Omega_G=0 in every G!=G0: otherwise all three of its arrays would need
an alternating direction. Equations (4) then imply G0=M and
Omega_M=-2. The two alternating directions are isolated at the two edges
of M other than U, with coefficient -2 each.

Relabel the entire configuration once as

    M={01,23,45},       U={01}.

This simultaneously fixes the D block; no separate symmetry assumption
is made on the star or glue data. In the fixed dictionary its doubled
column is 1437. Permit all stars, all glues, all N directions attached at
0 or 1, and even all eight group-M alternating directions whose isolated
pair is 23 or 45. All other retained types are zero by (9).

The following combination of **zero-based cell-row IDs** from the fixed
720-row cell-law file vanishes on every permitted column except 1437,
where its coefficient is -1:

    -1/2*(231+232+261+262)
    +1/2*(233+263+276+277+279+280+306+307+309+310)
    -1/2*(390+391+420+421)
    -(413+443)+509.

These are six T rows, fourteen V rows and one g row. The result JSON
records their full named keys, rational coefficients, the complete
combined row, and all 398 permitted column IDs. The replay recomputes the
combination directly from the literal rows and verifies that its
restriction is exactly `-D_1437=0`. This contradicts a=1.

The certificate is stronger than needed: it permits arbitrary support
and arbitrary signed coefficients on all allowed private columns, and
all eight permitted alternating columns. The proof does not depend on
the numerical behavior of a solver or on sampling coefficient values.

## 9. Coverage and global conclusion

Every normalized matching identity of retained support at most 23 has
been excluded:

1. N=0: the zero-zeta argument in section 4 gives matching retained support at least 24.
2. N>0, all zeta=0: section 4.
3. At least three N-active attachment vertices: sections 2–4.
4. One N-active vertex: section 5.
5. Two N-active vertices with at least two helpers, or a sole Z: section 6.
6. Two N-active vertices with a sole D: sections 7–8.

All counts are support counts of disjoint direction types. In particular,
Z is charged in Pi, and the loop cost is C=L3+E; the argument adds 2b
only when the charged flow blocks contain no Z. Mixed support is always
nonnegative, and is forced to zero in the final exceptional budget.

Combined with the fixed nonmatching bound, this gives the claimed
**global lower bound 27**. No upper certificate is changed. The bound
does not prove that 27 is attained or that 36 is optimal.



## Appendix A. The private lemma H=2M implies P>=24

This lemma concerns effective private stars and glues only. It allows
arbitrary signed rational coefficients and does not assume that the
private data extend to a full PM2 identity.

Index the ten private 3|3 partitions by sigma. Let a_(sigma,u) be the
center mass on u's private line, so a_(sigma,u)=-2s_(sigma,u).
Every vertex has total center mass one. Define

    B_uv = sum_(sigma: u,v in the same half) a_(sigma,u),
    H_uv = (B_uv+B_vu)/2.

Each B row sums to two. Each column also sums to two: in every private
3 by 3 glue matrix, the two half-sums of its center masses agree, so
the incoming sum at v equals total center mass/2 minus center mass at v,
namely 6/2-1=2. These identities are valid for signed marginals; the
incoming identities are additionally checked in all 90 glue coordinates.

Call u simple if exactly one a_(sigma,u) is nonzero. That marginal is
one, so its B row has two entries one and all others zero. If the two
endpoints of a matching edge are both simple, then H on that edge is
0, 1/2 or 1, and cannot equal two. Thus H=2M requires a nonsimple
endpoint on each of the three disjoint matching edges. At most three
vertices are simple, and at least nine private lines have nonzero sum.

For a line with glue vector l and s=sum(l), its effective stars are
l-s*1 and its cost is f=||l||_0+2||l-s*1||_0. Each glue is counted
twice and each star twice, giving 2P=sum f. If s is nonzero, the
possible costs are 5,7,8,9: with one nonzero glue entry the cost is five;
with two, it is eight; with three, the possible star counts are one,
two or three. If s is zero, the costs are 0,6,9. Nine nonzero-sum
lines therefore cost at least 45. The total is even, and an excess
of one over 45 is impossible from these spectra; the next possible
even total is at least 48. Ten or more nonzero-sum lines cost at least
50. In either case P>=24. This proves the lemma used in section 4.

## Appendix B. Literal restricted equations for the last case

Write x_j for the scaled coefficient of fixed canonical column j.
Each displayed equation is obtained from its literal cell row by keeping
the 398 permitted free columns and D_1437, and setting every other
retained coefficient to zero. The cell equations are homogeneous;
the correction columns do not occur in them. The multiplier column gives
the exact dual combination from section 8. There are six T rows, fourteen
V rows and one g row. Twelve V rows are centered at 2 or 3; the other
two are 413 and 443. The twenty-one rows consist of six T rows, fourteen V rows, and one g row.

| Cell ID | Kind | Multiplier | Restricted equation |
|---:|:---:|---:|---|
| 231 | T | -1/2 | `x_657 +x_965=0` |
| 232 | T | -1/2 | `x_171 +x_193=0` |
| 233 | T | 1/2 | `x_1281 +x_1286=0` |
| 261 | T | -1/2 | `x_678 +x_986=0` |
| 262 | T | -1/2 | `x_411 +x_433=0` |
| 263 | T | 1/2 | `x_1290 +x_1296=0` |
| 276 | V | 1/2 | `x_395 +x_986 +x_1369=0` |
| 277 | V | 1/2 | `x_396 +x_678 +x_1370=0` |
| 279 | V | 1/2 | `x_137 +x_965 +x_1362=0` |
| 280 | V | 1/2 | `x_138 +x_657 +x_1363=0` |
| 306 | V | 1/2 | `x_395 +x_433 +x_1408=0` |
| 307 | V | 1/2 | `x_396 +x_411 +x_1409=0` |
| 309 | V | 1/2 | `x_137 +x_193 +x_1398=0` |
| 310 | V | 1/2 | `x_138 +x_171 +x_1399=0` |
| 390 | V | -1/2 | `x_1296 +x_1369 +x_1408=0` |
| 391 | V | -1/2 | `x_1290 +x_1370 +x_1409=0` |
| 413 | V | -1 | `x_137 +x_138 +x_139=0` |
| 420 | V | -1/2 | `x_1281 +x_1362 +x_1398=0` |
| 421 | V | -1/2 | `x_1286 +x_1363 +x_1399=0` |
| 443 | V | -1 | `x_395 +x_396 +x_397=0` |
| 509 | g | 1 | `x_139 +x_397 -x_1437=0` |

The sum of these equations with the indicated multipliers is
`-x_1437=0`, contradicting its required value one. The complete named
keys, full untruncated rows and allowed-column list are stored in the
source result; the integrated replay checks the restriction afresh.
