Implementations

Six, all held to the same normative vectors (Appendix C).

./conformance.sh          run the 136 vectors against all six
python3 difftest.py N     feed N generated inputs to all six, compare byte-for-byte

language	location	NFKC from	notes
Python	coia.py	unicodedata	folds with str.casefold(), which is definitionally Default Case Folding; serves as the check on the generated table
JavaScript	impl/js/	String.normalize	ESM
Go	impl/go/	golang.org/x/text/unicode/norm	the only dependency; stdlib has no NFKC
Java	impl/java/	java.text.Normalizer	no external dependency
Rust	impl/rust/	unicode-normalization	plus unicode-properties; std has neither NFKC nor General_Category
Swift	impl/swift/	Foundation	precomposedStringWithCompatibilityMapping

Generated files

gentables.py builds tables.json from the rules in §5. gencode.py emits the tables, the case-folding table and the vectors as source for each language, so no implementation needs a JSON parser. Regenerate rather than hand-edit anything named data.* or Data.*.

Why the differential test exists

Vectors are fixed and were written by someone who already knew what the answers should be. The differential test feeds the same generated inputs to all six and compares byte-for-byte; it found a divergence at 3,000 inputs that the 132 vectors of the day and 300 differential inputs both missed — Go’s strings.ToLower applies simple case mapping, so İ folded to i there and to i + U+0307 everywhere else. That is what moved case folding from “lowercase, then patch the differences” to a complete normative table.

Current status: 136/136 vectors in all six, and 25,000 differential inputs with no divergence.

Known gaps

• coia.rs in the repository root is the COIA 1.x implementation and cannot run — the Rust regex crate rejects \p{Cs}, which its committed pattern contains. The v2 Rust implementation here is a rewrite, not a port.
• Language templates other than English are provisional pending native review; see Appendix A.