Appendix C — Golden vectors

Normative. Where this specification’s prose and these vectors disagree, that is a defect; report it.

The machine-readable set is vectors.json. Each entry carries the section it tests, so a failing vector points at the rule it came from.

Total: 136 — 69 normalize, 26 generate, 11 parse, 17 match, 10 reject, 3 search.

Vectors are authored from the prose, not captured from an implementation’s output. An implementation is not the oracle — that is how COIA 1.x’s specification and its Appendix A drifted apart without either being anyone’s bug.

Sections

section	count	tests
normalize	69	§5, input → canonical form
generate	22	§4, §6.1, (lang, who, role, scope, flags, private) → alias
reject	10	§4.3, §6.3, inputs a generator MUST refuse
parse	11	§6.1, §6.2, alias → (body, group1, group2)
match	17	§7, (alias, query) → boolean
search	3	§7, (corpus, query) → ordered list

null in a generate vector’s who position means the reflexive sentinel (§4.1), not an empty string — the two are required to behave differently.

Running them

python run_vectors.py

Exits non-zero on any failure, printing the section, the vector’s label, and both values.

Coverage intent

Every MUST in §4 through §7 has at least one vector. The normalize section additionally covers, by name: each SPLIT subclass (whitespace, dashes, quotes, script terminators, fraction slash); marks in Yoruba, Khmer, Dhivehi, Devanagari, Tamil and Hebrew; modifier letters in Japanese, Hawaiian, N’Ko and Lisu; all four joiner contexts in §5.1; NFKC folding of fullwidth, circled, ligature, superscript, Roman-numeral and halfwidth forms; and the degenerate inputs (empty, wholly elided, wholly separators).