The list ('based', 'class', 'type', ...) was seeded from two observed
cases and grew by intuition: English-only in an otherwise script-agnostic
design, unfalsifiable membership, and able to invent hyphens
('proto- type' -> 'proto-type'). It was added to prevent corruption under
the old unconditional default, and outlived the default it patched — under
evidence-only joining its entire corpus footprint is 8 cosmetic joins
across 7 documents, all of which now revert to honest visible breaks (the
documented no-evidence floor).
Every remaining rule is document evidence or script-agnostic typography.
A real hyphenation dictionary (e.g. TeX patterns) as an opt-in feature is
the principled path if those cases ever matter.
Abimael Martell
ed830a9e6a