Corrections in this revision

The supplied summary says “17 checks,” but its manifest contains 18: sixteen identity/normalization predicates and two norm-comparison predicates. All 18 are reported passed.

A rectangular isometry is not a unitary on its output space. The complete split isometry has orthonormal columns; it is not a square unitary matrix.

A floating-point residual printed as zero is not exact-arithmetic verification. The reported small residuals are retained as numerical evidence only.

The source-resolvent argument is made first for nonreal spectral parameters. The uncoupled positive-frequency continuum outside the interaction band is not incorrectly declared to belong to the resolvent set.

The autonomous chain represents the coupled cyclic subspace, not all uncoupled field modes.

The previously undefined preparation macro J is defined.

The geometric-tail quantity computed by the script is a majorant of the exponential-series tail, not the tail's exact sum.

A complete packet distance computed as $\sqrt{1-h^2}$ suffers cancellation when $h$ is near one. An algebraically equivalent stable formula is supplied. The archived execution values are not retroactively changed.

An exact truncation bound below floating-point roundoff is not a certified error bound for the computed matrices. A separate arithmetic allowance is required for that conclusion.

Smaller proxy survival than a continuum upper bound is consistency, not evidence that finite truncation always lowers survival. No such monotonicity is asserted.

The reported $T=1$ run is not the $T=40$ analytical small-error example. The former has substantial source survival and does not demonstrate recovery error below $1/100$.

Historical “unexecuted” wording in the archived version-1 script is provenance. Its reported execution status is described in Section sec:repro.