Corrections and exact mathematical release

The substantive debugging corrections are:

1. The identifier is consistently XVIII-A, version 2; the separate XVII-A benchmark is not overwritten.

2. Readout, observable representation, quantum return, and physical recovery are kept distinct.

3. Smooth witness non-determination is explicit; the singular example is used for failure of wavefront determination.

4. Class difference is not promoted to class admission.

5. Exact Ward cancellation uses the actual counterterm image, not its closure.

6. Finite-probe injectivity is stated on its declared vector space, with the weaker affine-family qualification kept separate.

7. Positive formal quantum orders do not repair a nonzero classical coefficient.

8. Normalized retarded extraction accounts for division by $\hslash$ and any explicit coupling dependence of the maps.

9. Balanced forests are not treated as loop products, and weak convergence is not wavefront convergence.

10. Finite normal-order ambiguity is distinguished from finite-dimensional ambiguity; point support is distinguished from positive-dimensional support.

11. Observable matching retains preparation residuals and a nonzero perturbative remainder.

12. Documentary stopping rules remain version-local; neither catalogue analogies nor technical lifting theorems change an inherited admission status.