scroll- craft.

The same reveal-on-scroll and parallax patterns kept getting rebuilt per client site. The fix was to write it once, correctly, and publish it, so every future build starts from a tested library instead of a copy-pasted IntersectionObserver.

Character and word-level text reveal effects work by wrapping every letter in its own <span>. Done naively, that means a screen reader encounters a wall of single-letter nodes and spells the word out letter by letter instead of reading it. A library that does this for a living has to solve that, not ship it as a known issue.

The library's rule is IntersectionObserver decides whether to animate, requestAnimationFrame runs the animation, and ScrollTimeline is never used, since browser support for it still isn't reliable enough to build on. Parallax and scene effects split reading layout from writing it: a measure() pass and a write() pass are kept separate and coalesced into one rAF tick (src/effects.ts:1040-1124), so the browser never has to interleave a forced layout read with a style write and thrash.

Every effect feature-detects the APIs it needs and falls through to the final visual state if they're missing, so a browser without some capability shows the finished layout instead of invisible content stuck mid-animation.

The accessibility fix: the text splitter sets a synthetic aria-label carrying the original full string, marks every generated span aria-hidden="true", keeps whitespace as real text nodes rather than spans, and removes the synthetic label again on revert so the DOM returns to its original state (src/effects.ts:384-473). A screen reader reads the label, not the exploded spans.