perf: mdx build time

[MohamedH1998]

Outcome: closed experiment — no measurable win

Build-time perf experiment for large sites (measured on Cloudflare Docs, ~8.7k
pages). All framework-only (packages/nimbus-docs). Conclusion: correct and safe,
but no measurable build-time win on the current stack — not shipping.

Two MDX-compile changes were tried here:

1. Skip the MDX partial-heading parse on Render-free pages — mergePartialHeadings
   only injects headings from <Render> slots, so a body without one is a pure
   pass-through; the mdxToMdast parse is skipped for it (byte-identical output,
   ~4,900 of ~6,700 cf-docs pages).
2. Opt-in persistent MDX compile cache (mdxCache option, default off) — disk
   cache around the @mdx-js/rolldown transform; a hit skips the compile.

A CI/A-B check showed no measurable change: this branch built in 367s vs 365/374s on
unrelated PRs (within noise). cf-docs's build is bound by bundling + rendering, not
the MDX compile these changes touch, so skipping/caching that work is a rounding
error. Same story as the sibling prerender-hashing experiment (#57), whose original
win also evaporated once Rolldown sped up its chunk hashing.

Disposition: closing unmerged. The two changesets on this branch never reach
main, so there's no release impact. Kept here as a documented negative result.

[pkg-pr-new]

pnpm add https://pkg.pr.new/@cloudflare/create-nimbus-docs@62

pnpm add https://pkg.pr.new/@cloudflare/nimbus-docs@62

commit: b4afebe

[MohamedH1998]

CI shows no measurable change: this branch built in 367s vs 365/374s on unrelated PRs (within noise). cf-docs's build is bound by bundling + rendering, not the MDX compile these changes touch - so skipping/caching that work is a rounding error.