Tested Cipher Outputs Under a Corpus-Induced Voynich Alphabet-Distribution Scorer

This successor narrows the result from class closure to test-set discrimination. Across the retained v3 and v4 trials, the unciphered manuscript surface scores above the tested catalog outputs under a scorer induced from that same manuscript corpus. Uniform-over-chip proxy outputs cannot reproduce the manuscript's non-uniform opener distribution. That is an alphabet-distribution mismatch, not proof against every constructed cipher: biased encodings, distribution-preserving transforms, alternative tokenization, post-processing, and methods outside the implemented catalog remain open. The modern per-primitive rows are relabeled seed variants of one SHA256-CTR proxy, not executions of AES, ChaCha20, Kyber, NTRU, McEliece, BIKE, HQC, SPHINCS+, or Dilithium. They support a common-proxy result only. Signature-scheme rows are structural simulations and are excluded from encryption claims. The Cardano-grille length argument is conditional on the paper's unsegmented multi-cell token grammar; an alternative emission grammar remains open. The Block B audit result is 12 of 13 folios at strength 1.0, with f99v.27 below the sampling floor; the earlier 8-of-13 statement is superseded. The corpus-induced scorer is partly self-fit, and independent medieval-corpus controls remain necessary.