Ship a build instead of sources, and keep the server renderer out of client bundles

[dave-hillier]

Closes #140.

The breakage

The runtime conditions resolved to src/**, so a consumer's bundler was handed raw .ts/.tsx. Vite therefore did not treat replot as an optimisable dependency, and its CommonJS dependencies reached the browser unconverted. A stock Vite app rendered nothing in npm run dev:

The requested module '/node_modules/react-dom/server.browser.js' does not provide an export named 'renderToStaticMarkup'
The requested module '/node_modules/interval-tree-1d/interval-tree.js' does not provide an export named 'default'

Production builds were unaffected, which is exactly why it survived CI and would have hit every new user on their first npm run dev.

Shipping a build

tsc -p tsconfig.esm.json emits an ESM tree into dist/esm; the runtime conditions point there and src is no longer published (the tarball drops from 326 files to 290).

tsc rather than a bundler: the sources already import each other with .js specifiers, so a structure-preserving emit resolves correctly as-is, and every dependency stays external without an allowlist to maintain.

One wrinkle worth knowing about — tsc silently declines to emit a .js source that has a hand-written .d.ts sibling, treating the declaration as the authoritative module. That is 30 of 47 files here, and the result was a dist/esm full of dangling imports. scripts/copy-sources.mjs copies those, mirroring the existing copy-declarations.mjs. Verified afterwards by resolving every relative specifier in the emitted tree: 141 modules, 0 dangling.

Keeping react-dom/server out of client bundles

Two separate paths pulled it in:

1. renderTransform.ts imported it to serialize JSX to DOM for the imperative render option. It now builds nodes directly — the inverse of domToJsx. renderJSX only ever emits intrinsic elements and fragments, so no renderer is required; a component element throws rather than emitting something subtly wrong.
2. plot.ts held both computePlot (imported by <Plot>, so every consumer) and the imperative plot() (which genuinely needs a renderer). react-dom declares no sideEffects: false, so a bundler keeps the react-dom/server import of any module it pulls in, even when the binding is unused — tree-shaking could not save us. plot() moves to plotDom.ts, leaving computePlot clean.

Measured on a real consumer

Tarball installed into the example app, with its optimizeDeps workaround removed:

	before	after
dev server	fails, 0 charts	11 charts, 0 console errors
client bundle	788 kB / 256 kB gzip	601 kB / 197 kB gzip
react-dom/server modules bundled	3	0

Production preview also renders 11 charts with 0 errors, and strict tsc -b in the consumer stays clean.

Tests

The render-transform path had no coverage at all, so the walker would have been an unverified rewrite. test/react-render-transform-test.tsx adds four tests over the DOM next() hands back — namespace, camelCase-to-hyphenated attributes, className to class, and the output reaching the document. Mutating the attribute mapping fails one of them, so they are not vacuous.

yarn test:mocha 1567 passing (up from 1563), tsc/lint/prettier clean, publint --strict all good, attw green for both entries.

Follow-up

replot-vite-example still carries the optimizeDeps.include workaround and documents it. It should be removed once a version containing this fix is published — doing it now would break the example against the currently published 0.1.0-alpha.0.

🤖 Generated with Claude Code