M14-v1 independent review and exact replay.
The saved cover is traversed from the whole rational time cube using the
declared bisection rule. All28097 nodes terminate at exactly the14049 saved
leaves or an exactly infeasible simplex box; there is no uncovered feasible
cell. Every leaf's lower bound is then recomputed using the independent row
LP dual rather than the greedy primal:
 l.v + eta(m-sum l) -sum_j(u_j-l_j) max(eta-v_j,0).
Selecting eta at a capacity threshold gives equality. All leaf numerators
match exactly, including the minimum2593001527986953 over denominator10^16.
This shares the validated spectral envelopes, so it is independent at the
cover and linear-program layers only, not an independent numerical library.

Arithmetic audit: exact residual enclosures use the two nonzero off-diagonals
of H and integer-quantized eigenvectors. Orthogonality uses integer products,
not a solver-reported error. The proposed spectral matrices need not themselves
be exactly orthogonal; the polar comparison explicitly charges that defect.
The stated kernel budgets dominate gamma_(8n) bounds for n<=36 and J,A<1000.
The A<1000 assumption is also added as an explicit executable assertion in
candidate v2. All kernel and memory thermal weights pass exact interval checks.

The reset lower bound discards a positive initial subensemble. It does not
normalize it and cannot be biased upward by missing thermal tails. Its row
mass is rounded DOWN, and its backward value grid is rounded DOWN. High-cost
missing probability can only lower this relaxed objective. The memory upper
bound includes its own positive missing-sector energy and arithmetic margin.
The controls allow different selected machines and phases on the reset side:
fresh replacement makes their target marginal channel identical and phase
independent. At most four pulses includes the witness's three and all zero
time reset slots. There is no resource mismatch.

The original PRX source2512.10703v2 explicitly leaves memory effects of the
non-Gaussian p-excitation exchange for future study. Its Gaussian no-memory
advantage and asymptotic reset results do not cover this finite-budget witness.
No global novelty or external review claim follows. Internally supported
complete-original candidate, with the IEEE/MPFR error contract stated in the
proof. An independent interval-library implementation would be a useful
external replication, not something silently claimed as already performed.
