QL: Lorentz join normalizes an exact-zero propagator column

[krystophny]

Problem

The Lorentz flux-conservation correction in NEO-2-QL/join_ripples.f90 divides every joined propagator column by its column sum without handling an exact-zero column.

At current stacked head 03f5b9dacdbf7c882fac01c37972e8e266575351, a Debug build with bounds checking and -ffpe-trap=invalid reaches:

join_ripples.f90:294
facnorm = 0
amat_p_p(1,2) = 0
amat_p_m(1,2) = 0
SIGFPE: 0/0

Context:

i = 2
npass_l = 24
npass_r = 24
old propagator tags = 12..13
new propagator tags = 13..13

The collision operator and propagators complete before this fault. No earlier NaN, IEEE exception, or failed linear-solve diagnostic appears in the trapped run.

Expected behavior

The flux-conservation correction must define and verify the intended behavior for a zero incoming propagator column. It must not silently create NaNs or divide by zero. Nonzero columns must retain the existing normalization and conservation law.

Acceptance

• Add a behavioral fixture that reaches a joined exact-zero Lorentz column and fails before the repair.
• Define the physical meaning of the zero column: impossible state, removable empty channel, or upstream assembly defect.
• Preserve column-sum-one normalization for every nonzero column.
• Reject an impossible zero column with stage, index, dimensions, tags, and finite-value diagnostics, or handle a physically empty channel with an explicitly justified invariant.
• Run the full native suite and a bounds-checked source-only helical-geometry case.
• Verify no IEEE invalid flag and finite final transport outputs.

[krystophny]

The upstream trace identifies the zero-column producer at the current stack:

• ripple_solver_ArnoldiOrder2_test.f90 allocates all four output transfer matrices at lines 827-834 but does not initialize or construct them afterward.
• The HELCORE path selects ArnoldiO2 (isw_ripple_solver=3) and requests non-axisymmetric qflux, so joining consumes undefined child transfer matrices.
• Column 2 is the second m=0 passing pitch channel: the Lorentz correction loops over 1:npass_l, not source or higher-velocity slots. With npass=24, this is a real channel.
• With bin_split_mode=0, the join converters are identities. Conservative child transfer maps must preserve a nonzero outgoing sum through composition; an exact zero is not a roundoff normalization error.

The required producer-side fixture is one ArnoldiO2 field propagator before its first join. It should assert that all four matrices are finite and that each Lorentz density-channel outgoing sum is nonzero. For every accepted joined column, the existing conservation invariant remains:

sum(amat_p_p(:,i)) + sum(amat_p_m(:,i)) = 1
sum(amat_m_p(:,i)) + sum(amat_m_m(:,i)) = 1

PR #135 prevents the downstream 0/0 and reports the exact channel, but deliberately does not close this issue or invent transfer values. The remaining numerical-design decision is whether ArnoldiO2 must construct boundary-response transfer maps or its global-solve mode must bypass propagator joining through a separately justified interface.