Skip to content
fold-kit-experiments

Composable FoldKit charts qualification

Date: 2026-10-04. Checkout: managed codex/foldkit-viz-promo worktree.

Automated evidence

  • Root bun run check: exit 0, no lint or format findings.
  • Root bun typecheck: exit 0, Astro diagnostics contain zero errors/warnings/hints.
  • Root bun run test: 384 passing tests (Viz 168, integration 74, promo 37, web 105).
  • Promo production build: exit 0, six pages generated.
  • git diff --check: exit 0.
  • Independent OS temporary projects for line, histogram and scatter: npm install, typecheck and build all exit 0; all three new public subpaths import under Node.
  • Packed package: root/geometry/theme work under Bun and Node; real npm pure installation leaves FoldKit and Effect absent; an independent installation of both optional peers renders adapter output under Bun and Node.

Browser evidence

The in-app browser loaded the actual dev server at 127.0.0.1:4321. Each of line, histogram and scatter was checked at 390 and 1280 CSS pixels, in both light and dark themes (12 combinations). Document scroll width equalled viewport width in every combination. Computed axis labels were 12px throughout. Measured chart width was 324px on mobile and 818px on desktop; viewBox width followed it.

Line controls updated the chart description, raw table and maintained settings source together (first observation 10 → 11). Two series, a keyed legend and the 80-unit annotation remained visible. Histogram dataset/bin controls changed the bar count from five to six, source agreed, and the raw table exposed 40 records. Reset restored five spread bins. Scatter End inspected b-12 with original X=95, Y=8; selection and the custom tooltip survived a resize to 390px. Pointer inspection worked, filtering to the opposite group cleared selection, reset restored 24 marks, and its raw table had 24 records. Circles and squares matched the legend and retained group paint.

Native browser input changed the StackBlitz editor buffer. A session marker survived tab switches, theme change and mobile/desktop resize. Restart created a fresh editor with the current first observation of 11. WebContainer installation completed and its nested preview rendered the composed chart. Editing chart.css to set --chart-series-a: oklch(65% .15 150) updated the real nested preview’s computed stroke to oklch(0.65 0.15 150), with the reference stroke retained.

Download produced visible success feedback. The resulting native Downloads ZIP contained 43 files, shared frame/measurement sources and settings with first observation 11, agreeing with the controls/source rather than editor edits. Histogram and scatter downloads contained 42 and 41 files respectively, with clustered/five-bin and inspected a-02 settings matching their controls and source.

Evidence boundaries

Changing initialSettings in a running editor preserves the existing runtime Model; the preview does not reset live control state merely because defaults change. Reload the preview to apply edited defaults. CSS edits update directly. The browser download-event waiter did not observe the native download; the success state and actual ZIP contents confirmed it independently.

No site deployment, npm publication, release version bump or full-catalogue migration was performed. Mathematical colour interpolation remains hex-only; the existing RGB basis interpolation parity gap is outside this slice.

Fresh review and fix pass

A fresh gpt-6-astra reviewer inspected the whole branch, 4bd5016..a9457c6, read-only. No Critical finding; three Important numeric findings were reproduced: finite projected coordinates causing non-finite spline output, collapsed epoch millisecond ticks, and unrepresentable histogram thresholds producing duplicate interval keys. Its Minor README type issue was regraded to Important because copying the example failed compilation.

All four regressions were observed failing first (23 pass, four fail), then passing (27 pass, zero fail). The geometry boundary now rejects non-finite curve paths and non-distinct generated thresholds. The tick implementation follows pinned d3-array integer/reciprocal increments. The primary README snippet is compiled against the packed optional adapter. All material findings were fixed; no deferred minor remains. Full verification repeated after the fixes.

Rulings retained from implementation and review:

  1. Name the pure local factory linkedChartFolds, since the service lint treats make-prefixed imports as Effect services. Cost: one internal name differs from the plan; public API is unchanged.
  2. Consume scatter semantic notifications through foldChild in its Storybook host. Cost: the story has no sibling coordination, while Commands survive.
  3. Run workspace tests sequentially because package smoke rebuilds Viz output that promo packaging reads. Cost: modestly longer tests.
  4. Keep comprehensive existing curve/colour parity outside the slice, while validating new accepted geometry. Cost: other primitive parity gaps remain.
  5. Keep publication/deployment outside the slice. Cost: external availability requires a subsequent release/deployment.
  6. Keep full-catalogue migration and large-dataset performance outside the three-example scope. Cost: other charts retain earlier patterns and large datasets remain unqualified.