[elementmerc]

Follow-up to #2. Switching the example dropdown on the live site took about 1.2 seconds; this brings a warmed switch down to roughly 150ms.

Cause

Two network costs paid at click time:

• loadPreset fetched each .alt with { cache: "no-cache" }, forcing a network round trip on every switch.
• The first switch to a chart or diagram article downloaded its graphics library (uPlot, KaTeX, Mermaid) at that moment.

Fix

• Cache preset sources in memory; a repeat switch renders synchronously with no network.
• On idle (requestIdleCallback), prefetch every preset source and warm the lazy graphics libraries, so the first switch to any example is instant too. The runtime imports the same specifiers, so the browser serves them from its module cache.
• Drop a late fetch result if the user has already switched away.

Measured (live, before vs target)

Before, per switch on the deployed site: cybersecurity 1157ms, cs-lewis 1228ms, paradox-of-genius 1251ms, languages 1037ms. Warmed local (no network) is ~115 to 215ms; the live warmed numbers will be posted on this PR once deployed. No crate changes; playground-only.

[elementmerc]

Live measured, deployed, milliseconds per switch:

Example	Before	After (cold)	After (warmed)
cybersecurity	1157	119	178
cs-lewis	1228	65	83
paradox-of-genius	1251	61	50
languages	1037	818	178

A real switch (idle warmUp completes while the visitor reads) lands around 50 to 180ms, down from roughly 1.2 seconds. The cold "languages" outlier is the case where the visitor switches before the idle prefetch finishes and several charts render at once; still well under the old number.