BPE Morpheme-Length Clustering Across 55 Writing System Corpora: Typological Validation and the Voynich Discriminant Zone

Paper 6 (Silva 2026) introduced BPE Mean Vocabulary Morpheme Length (VMML) as a writing system classifier and showed that the Voynich Manuscript occupies a discriminant zone (VMML = 5.918, 95% CI 5.77–6.05) above all 15 tested alphabetic natural languages. Here we expand the comparison to 55 deduplicated corpora spanning 25+ languages, 5 morphological types, and multiple script families, and add three within-token boundary metrics (BC, BPE-ratio, CBMI) computed across all corpora. Three findings extend Paper 6: (1) The alphabetic ceiling is confirmed at 5.76 across 40+ large-corpus controls (n ≥ 10,000 tokens), widening the validated reference set by 3×; (2) Tagalog (Austronesian agglutinative, n=173,885, VMML=5.914) is the sole natural-language corpus entering the Voynich 95% CI — but its internal boundary concentration (BC=0.202) is the lowest of all 55 corpora, sharply distinguishing it from Voynich (BC=0.361) despite identical VMML; (3) Cross-boundary mutual information (CBMI) clearly separates random gibberish (0.351 bits) from all tested natural languages (0.486–1.125 bits), with Voynich (0.710 bits) occupying an intermediate zone consistent with a constrained phonotactic inventory. Romanized transcriptions of non-alphabetic scripts produce systematic VMML inflation (Δ = 2.4–5.3 units), constituting a methodological confound excluded from the primary analysis. After applying corpus-size (n ≥ 10,000) and script-type controls, the three-cluster structure of Paper 6 is confirmed and refined: no tested hoax-generation mechanism enters the Voynich CI on VMML, and no corpus replicates Voynich's combination of high VMML and edge-concentrated boundary distribution. All code and data are publicly available.