[JLHwung]

In this PR we shrink the output size by more than 50% with the help of LLM:

Reduces the published @unicode/unicode-18.0.0 packages from 700 KB to 312 KB packed, from 1.96 MB to 1.08 MB unpacked, and from 8,344 to 4,320 files. No module path, export shape, regex source, or map entry order changes; every step was verified against the previously generated output for all 35 Unicode versions.

Changes

Commit	Change	Packed size after
bfbcb5f	Compress inline data with brotli instead of gzip	590 KB
7fb6a67	Build regex.mjs at runtime from ranges.mjs	444 KB
4b18111	Derive mapping symbols.mjs from code-points.mjs	410 KB
86b542e	Front-code the Names table	390 KB
0086ab8	Delta-encode case and decomposition mappings	378 KB
0506410	Share type declarations via typesVersions	312 KB
22387c7	Declare the default export of the package index	—

Brotli. gzipInline becomes brotliInline at maximum quality and window size. Generated modules import brotliDecompressSync from node:zlib.

Runtime regexes. Each regex.mjs held a regex literal from Regenerate; astral ranges expand into long surrogate-pair alternations (470 KB per package). regex.mjs now imports ranges.mjs and calls the new ranges-to-regex.mjs, a port of Regenerate's serializer. source and flags are byte-identical to the previous literals.

Derived symbols. For the four mapping types, symbols.mjs carried a second compressed copy of code-points.mjs. It now derives the map at runtime through mapping-to-symbols.mjs. Mappings with fewer than 10 entries keep their inline literals.

Front-coded Names. Names/index.mjs stores one name per line, front-coded against the previous name, with the run pairs as a JSON first line. decode-names.mjs rebuilds the same Map<number, string>. 158 KB → 123 KB.

Delta-coded mappings. Case and decomposition code-points.mjs files store zigzag deltas as variable-length integers in the same base64 scheme as ranges.mjs, decoded by decode-mapping.mjs. A count of 0 marks a scalar value, so files that mix scalar and array values (Unicode 3.0.1 Case_Folding/E) round-trip exactly.

Shared declarations. The four identical per-directory .d.mts files are written once under types/ and selected through a generated typesVersions map. Mapping directories keep their own code-points/symbols declarations because their value types differ; more specific patterns route them there. The build removes stale per-directory declarations from earlier runs.

Fixes found along the way

• ranges.d.mts imported a UnicodeRange type that decode-ranges.d.mts only declared inside a namespace, so TypeScript consumers got an error on every ranges.mjs import. The type is now exported at module level.
• index.d.mts declared named exports while index.mjs only has a default export, so import dirs from '@unicode/unicode-X' did not type-check. It now declares the default export with one readonly <Type>: string[] per property type.

Things to be aware of

• Node.js floor. brotliDecompressSync requires Node.js 11.7 or later. The previous output needed only gunzipSync. Note that although we have not exactly specified the Node.js support, we already use Object.fromEntries in decode-ranges.mjs, which requires Node.js 12.0.0, so the switch from gzip to brotli should be safe.
• TypeScript floor. .d.mts files already required TypeScript 4.7; the generated README now says so explicitly. typesVersions itself has existed since TypeScript 3.1.
• Import cost. regex.mjs and mapping symbols.mjs now do a little work on import instead of shipping precomputed data. Names/index.mjs builds its map on import, as before.
• Publishing. All 35 output/unicode-* trees were regenerated locally and need to be committed and tagged in their own repositories as usual.

Verification

• npm test passes (21 tests). New tests assert that every regex.mjs matches Regenerate's output for the same ranges, that every mapping symbols.mjs matches its code-points.mjs, that representative Names lookups are correct, that the declaration layout and typesVersions targets are consistent, and that index.d.mts matches the keys exported by index.mjs.
• One-off checks against the previous output across all 35 versions: 19,610 regex sources identical, 1,699 mapping symbols.mjs maps identical, 1,699 mapping code-points.mjs maps identical (entry order included), and all 35 Names maps identical by hash.
• Type resolution checked with TypeScript 5.9 under node16, bundler and node10 resolution, and with TypeScript 4.7: imports of regex.mjs, ranges.mjs, code-points.mjs, symbols.mjs, a mapping code-points.mjs, Names/index.mjs and the package root resolve to the same types as before.

[JLHwung]

As for the Node.js floor, I suggest we explicitly specify the engines requirement in package.json, but that should be addressed in a different PR.

[mathiasbynens]

Thanks for your work on this!

I’m considering merging this except for the runtime regexes (7fb6a67). This is the only change that I struggle to get behind. I like that it’s easy to look up the regular expressions for a given property/value pair just by checking e.g. https://github.com/node-unicode/unicode-18.0.0/blob/main/Script_Extensions/Adlam/regex.mjs, and that was one of the original motivations of this project (as the README change in that comment indeed suggests). I don’t like that static/ranges-to-regex.mjs reimplements parts of Regenerate, either.

[mathiasbynens]

I’ve merged the changes except for the runtime regex patch. Thanks a lot!

As for the Node.js floor, I suggest we explicitly specify the engines requirement in package.json, but that should be addressed in a different PR.

Agreed.