One Successful Seed Forces Universality: Selection Jump and Witness-Shape Laws in Exact Witness Architectures

The central theorem of this paper is a selection jump. For a decidable source-independent stagewise verifier (n,c), let\[S_{}=\{e:∀ n\,∃ t\,((e,n,t) (n,(t)))\}.\]If S_≠, then S_ is Π⁰₂-complete. Thus a single successful seed already forces maximal stagewise complexity. We prove a companion one-shot calibration theorem: for every decidable $D(c)$, the class\[O_D=\{e:∃ t\,((e,0,t) D((t)))\}\]is either empty or Σ⁰₁-complete. Together these results yield a witness-shape universality principle for exact local verification: one-shot local exactness yields Σ⁰₁-universality, while stagewise local exactness yields Π⁰₂-universality. We place this principle inside a broader barrier hierarchy. At the lower level, finite stagewise prefixes are uniformly insufficient, and undecidable co-c.e.\ witness architectures admit neither decidable one-shot positive certifiers nor exact-domain witness compilers. At the higher level, same-theory adequacy along a universal Π₁ embedding yields full Π₁ reflection, while arithmetic exact terminality on a truth-faithful image exists exactly when the corresponding fragment truth set is arithmetical, yielding Tarski and diagonal barriers. For fixed propositions with exact two-sided decidable witness packages, these results assemble into a two-sided bridge trichotomy: isolated extensional bridges are vacuous; the effective bridge layer splits into a positive stagewise fibre that is empty or Π⁰₂-complete and a negative one-shot fibre that is empty or Σ⁰₁-complete; and assertion-enriched resolver layers are truth-universal.