fix(Link): restore prefetching under Nuxt 4.5's custom slot

[IgorShevchik]

Ports c4ee0ea3 and fbb9e220 — a contiguous pair taken in one PR per §6 4b; the second is n/a and carries no change. Cursor moves to fbb9e220, which is nuxt/ui v4 HEAD, so the sync is caught up.

The regression was real here, and read from the installed source

Nuxt 4.5 stopped prefetching custom links itself. Rather than trust the upstream subject, this was checked against nuxt-link.js in the nuxt@4.5.2 this fork declares:

line	guard
:225	if (import.meta.client && !props.custom) — the visibility observer
:297	if (!props.custom) { … shouldPrefetch('interaction') … } — hover and focus

Link.vue here always renders <NuxtLink … custom>, so neither fired. Nothing errors and nothing warns; navigations are simply slower than they were — the quietest kind of regression.

The same file exposes the replacement API to the slot at :284 (prefetched, shouldPrefetch), which is what makes this a re-wiring rather than a reimplementation. And :210 computes props.custom ? ref?.$el?.nextElementSibling : ref?.$el — the origin of the "root is a fragment anchor, the element is its next sibling" step in the port.

The port

src/runtime/utils/prefetch.ts is taken as upstream wrote it — a requestIdleCallback compat pair and one shared IntersectionObserver. Its header note holds here unchanged: kept out of utils/link.ts so the Vue builds, which share LinkBase and never render NuxtLink, do not load it.

Link.vue takes the wiring with the fork's standing renames only (b24ui, B24LinkBase). Our resolveLinkClass keeps its extra isAction variant, so prefetchedClass joins the existing class rather than replacing upstream's shorter call. Both template branches are wired, as upstream does.

A guard of ours that upstream does not have

jsdoc-coverage rejected the new module: upstream documents two of its three exports with line comments and the third not at all, while our rule (#89) wants a /** */ block over every export from a published path.

Writing them pinned down something the line comments left implicit — a missing handle in cancelIdleCallback is a no-op, so an unmount before the callback was ever scheduled needs no guard at the call site.

The tests can see this port

Worth saying explicitly, because #527 earlier in this same queue was the opposite case — there the ported handler is never invoked under happy-dom and its tests stay green with or without the fix.

These assert against the link:prefetch hook on the Nuxt app rather than against markup, and drive the IntersectionObserver by hand (happy-dom's is inert). Five cases: both triggers, the preservation of a caller's own pointerenter / focus handlers, and the two opt-outs.

Mutation-checked four ways, each red on a different subset:

mutation	red
remove both getPrefetchListeners spreads — revert the port	3 of 5
disable the visibility observer	1
drop the caller's listeners from mergeProps	1
ignore shouldPrefetch('interaction')	1

fbb9e220 — n/a

Four files, all CodeTree: the fix, 169 lines of test, a new 391-line playground page added as the surface for it, and one line registering that page.

src/runtime/components/prose/ here holds 43 components and CodeTree is not among them — checked by listing the directory rather than grepping the name, so a rename would have shown up. Nothing in src/theme/ references it either.

Recorded, not acted on: this fork has 43 prose components and no playground page for any of them, while upstream now has one. That is a gap of ours rather than a port — building it would mean writing our own page for our own components. Written into the log and left, the same shape as the InputMenu playground row found while backfilling 2799fa6f and fixed later in #520.

Verification

lint · typecheck · build (3.87 MB, up from 3.86 — the new module is visible in the size) · test — 343 files, 7847 passed, 6 skipped · test:module.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JS8ypVfQSFzYVZzkTHhURb

---

Generated by Claude Code