One persistent signal: the thread survives the route handoff
The falsifiable questionCan a site's core claim — intelligence survives the handoff — be made structurally true: one visual element that literally never unmounts across navigation?
External validation withheld: no one outside this site depends on the component yet.
Baseline
The previous implementation rendered 32 sealed SVG decals, each drawing once on scroll-into-view and freezing. A repo-wide grep for scroll-linked motion values (useScroll, scrollYProgress) returned zero hits, and the motif unmounted and redrew from nothing on every route change — the literal opposite of the thesis.
Method
Replaced the decals with a single dependency-free 3D canvas mounted once in _app.tsx, above the router: a perspective camera (position, look-at, focal length) with keyframed shots — dolly-zoom at the handoff, orbit at the memory knot, crane at resolve — driving one yarn through six topologies as a pure function of scroll position. Persistence was verified by tagging the canvas DOM node with a data attribute, navigating Home to /projects/orgx, and confirming the identical node was still mounted and drawing.
Measured result
Same canvas DOM node alive across the route change (verified by attribute survival, 12.46% lit pixels post-navigation). Render cost measured at 3.1–8ms per frame at 2880×1800 — 60fps with headroom. Chartreuse accent held to 0.2–0.9% of pixels against a brand budget of 8%.
Where it broke
Three defects on the record. (1) The first material pass used ctx.filter blur on long strokes, costing multiple seconds per frame — replaced with stacked low-alpha strokes at ~100x lower cost. (2) Variable-width chunked strokes rendered as capsule seams at high zoom — replaced with a single variable-width ribbon fill. (3) After the merge, the signal was invisible: .site-shell carried an opaque background that painted over the canvas site-wide. Found by walking the live site, fixed in a follow-up commit.
Updates
- Published. Merged as PR #24.
- Occlusion defect found on the live site (.site-shell opaque background) and fixed; failure section updated.
- Superseded. PR #31 replaced this canvas with the material knot, so the component no longer runs on the live site and the demo link no longer shows it. Its job moved to the page thread: one strand sewn through every page, carried across route changes.
Next test
Page-transition FLIP handoff: the lit node slides through the door into the case-study spine instead of easing in place.