Verify.

01

Generated Vue & Svelte files had an inert stroke.

The starter file from npx svg-scroll-draw init used JSX-style attribute names in HTML-parsed templates. Both graphics below request an identical 2.5px round stroke. Only one gets it.

02

getProgress() was frozen at zero.

For every autoplay stroke animation, currentAlpha was only assigned on the clip-path branch — so the public getProgress() API reported 0.000 forever. The readout below is polled every frame from the real instance.

03

Off-screen pause/resume burned the animation out.

Leaving the viewport set startTime = null. Because null coerces to 0, a later pause() recorded the whole timestamp since page load as elapsed time — so resume() started a run already past its own duration and completed instantly, invisibly.

idle — press the button

04

Every size figure is now measured.

The README claimed ~4.4 KB against a real 8.9 KB, directly beneath a bundlephobia badge showing the true number. These come from npm run size, which reads the built output and fails CI if any entry drifts past its budget.

9 of 21 entries shown — the full table lives in README.md.

05

Claims that were wrong, and now can't be.

scripts/check-claims.mjs derives the real test and example counts from source and fails the build when a doc disagrees. It caught a stale number mid-session, which is the only reason this table is right.

06

Native CSS and the JS engine disagreed.

The library's headline claim is that the native CSS fast path and the JS engine are interchangeable. Nothing verified it, and it was false. Below, the same SVG is drawn twice at the same scroll offset — left on whatever the browser supports, right with native: false forced. Scroll slowly and watch Δ.

07

The CDN build crashed in a plain browser.

Every dev warning in 13 modules was guarded by a bare process.env.NODE_ENV check. process does not exist in a browser without a bundler, so reaching a warning threw instead of logging one. Both guards are evaluated below in a scope that has no process.

08

The JS engine worked while the page sat still.

The rAF loop ran every frame for as long as the container was in view, whether or not the scroll position had moved — recomputing values that had not changed. This is the path Firefox and every pre-115 browser always take, so it is not a niche case.

09

scrollSnap overrode prefers-reduced-motion.

The clearest accessibility defect in the library, and the readout below is live against your own OS setting.

10

scrollHorizontal never moved the track.

The worst defect in three phases of this work. The default trigger window was measured against the track itself — and a sticky stage pins that track at exactly one stage tall, so both ends of the window resolved to the same scroll position. Progress was then clamped at zero forever. The API did nothing, in precisely the CSS setup its own docs prescribe.

11

destroy() left elements stranded mid-animation.

The engine wrote opacity and transform inline every frame and removed neither on teardown. Destroy a component while its reveal was half-finished and it stayed half-finished — permanently faded and offset, with nothing left running to correct it. Worse for the visitor than never animating at all.

12

onSnap fired twice for a single snap.

scrollSnap reacted to the scroll event its own animated scroll produced, treating it as a fresh user gesture and announcing a snap it had already announced. Guaranteed under reduced motion, intermittent otherwise — a public callback that fires once or twice depending on the easing curve.

13

A finished sequence reported 0%.

scrollDrawSequence delegates getProgress() to whichever step is active. When the last step completed, the active index advanced past the end of the array, so the getter read from an undefined instance and fell back to zero — a sequence that had visibly finished reported that nothing had happened.

14

split: 'lines' deleted the spaces between words.

The only one of the three you can see without instrumentation. splitIntoLines groups words by offsetTop, and the whitespace nodes between them were dropped rather than reinserted — so a split headline rendered as one unbroken word. The sentence below is the test: read it.

15

What still is not proven.

The limits, kept up to date rather than quietly dropped as they shrink.

• —CI going greenThe browser suite has only ever been proven on one laptop. Pushing the v2.10.0 tag triggered the first remote run; its result lives on GitHub Actions, not here. That run also fails deliberately at "is this version already published", because 2.10.0 went out by hand — see the next item.
• —2.10.0 has no provenance attestationThe release workflow publishes with --provenance, which cryptographically ties a package to the commit and workflow run that built it. It needs an NPM_TOKEN repository secret, which was not set, so this version was published from a laptop instead. The workflow is wired and waiting; the claim is simply not yet true of anything on npm.
• —531 unit tests are still jsdomgetTotalLength is stubbed to a constant and IntersectionObserver is faked. They verify engine arithmetic. The 175 Playwright tests per engine are what cover browser behaviour, and a separate suite with no DOM at all covers what happens on a server.
• —The wrappers are tested through their contracts, not their build pipelinesAll eight now mount for real and are held to the same contract — the engine runs, unmounting leaks no observer and no frame loop, re-mounting works. But Svelte is exercised through its action functions rather than a compiled component, and Angular through its ref classes rather than a running Angular app. That is the whole public surface in both cases, and it is still not the same as a real build.
• —Changing an option after mount does nothing in most wrappersOnly the Svelte actions re-create the engine when their parameters change, because Svelte calls update() for you. React, Vue, Solid and Angular read their options once on mount, so a changed prop never reaches the running engine. Defensible — re-creating on every render would thrash — but undocumented and surprising. The tests pin it so it cannot change by accident while the decision is open.
• —This page found a bug the e2e suite missedSection 06 kept reporting a constant 0.0201 offset. Trigger points were cached at init and only recomputed on a window resize, so a 21px layout shift during hydration left the JS engine permanently offset while the native CSS path stayed correct. Now fixed with a ResizeObserver — but only a busy, real page surfaced it, which says something about how much the isolated fixtures can prove.
• —Dev warnings still need opting intoThey were unreachable for CDN users entirely: IS_DEV is derived from process.env.NODE_ENV, and process does not exist in a browser without a bundler. 2.10.0 ships svg-scroll-draw.dev.global.js, which reports them — but you have to know to load it, and the production bundle has them removed at build time.
• —Two of the three 2.10.0 fixes are now demonstrated — sections 13 and 14A finished sequence reporting its real progress, and split: "lines" keeping its spaces, both now measure the current build live rather than replaying a recorded number.
• —The named-easing parity demo is withheld, and that is a finding in itselfA harness written for it measured the native CSS path and the JS engine 0.83 apart with easing: "ease-out", against 0.00 for the same harness on the default curve — one variable changed. That is far larger than the 0.069 divergence 2.10.0 fixed, and too large to be an easing curve difference, which points at the animation range rather than the timing function. It is one new harness on one browser and it may well be measuring wrong, so it is not being published as a verdict about the library. It is written up in NEXT-SESSION-PLAN.md with the repro, and this page will carry the section once the number is understood.