Empirically measured under real trace-based (devtools) CPU/network
throttling against the live site: mobile Performance 59, LCP 9.2s
(TTFB 745ms + 8.4s element render delay).

- sidebar-components.tsx / [lang]/layout.tsx: the docs sidebar renders
  every page in the doc tree as a link at once. Next's default
  viewport-prefetch fired an RSC payload fetch for every one of them on
  initial load - dozens of concurrent requests competing with the page's
  own content for bandwidth. Wired fumadocs' documented `sidebar.prefetch`
  option through to the custom SidebarItem/SidebarFolder components (which
  were bypassing it entirely, using next/link directly with no prefetch
  prop) via the `useSidebar()` context hook.
- video.tsx: `autoPlay` forces browsers to fetch the full video file
  immediately on mount regardless of `preload`. Gated actual src loading
  behind an IntersectionObserver so a page with several of these doesn't
  pull down every video up front (5MB across 3 requests, in this case).
  Single shared component - fixes every doc page that embeds one.
- proxy.ts: the i18n middleware matcher excluded favicon/robots.txt/etc
  but not `icon.svg`, so every request for it got routed through i18n
  negotiation instead of served as a static file, 404ing in production.
- next.config.ts: enable productionBrowserSourceMaps - safe since this
  repo's source is already fully public, real debuggability benefit,
  zero performance cost.
- shiki 4.0.0 -> 4.3.1 (verified: syntax highlighting still renders
  correctly). Attempted a coordinated fumadocs-core/ui/mdx/openapi
  upgrade to latest; fumadocs-openapi's v11 factory function became
  client-only (breaking change beyond its declared peer deps, requiring
  a component-boundary restructure), so only the safe, verified,
  docs-exclusive bumps (fumadocs-core/ui/mdx, shiki) are included here -
  the openapi major bump needs its own dedicated migration PR.

Verified via a real production build (dummy env, all 3974 pages
including API reference render/build cleanly) and a clean (non-stale)
local server: Performance 59 -> 71 measured under real devtools
throttling, RSC prefetch requests 63 -> 11, video requests/bytes 3/5MB
-> 0. A pre-existing React hydration warning (#418) was found and
confirmed present on live production before any of these changes,
unrelated to this diff - documented, not blocking.

The lazy-load gate from the previous commit threw before isInView could
ever become true in environments lacking IntersectionObserver (older
browsers, some embedded webviews), leaving videos permanently
source-less instead of falling back to eager loading.

Repo convention is TSDoc-only, no plain // comments.

Audited with parallel subagents against the accessibility and SEO skill
checklists, each fix verified by reading the actual code (not assumed):

Accessibility:
- lightbox.tsx: focus was never captured/restored on close, and Tab
  escaped the modal to the page behind it (no focus trap on the single
  focusable element)
- heading.tsx: the per-heading copy-link icon only appeared on hover,
  invisible to keyboard-only navigation (added peer-focus-visible)
- navbar.tsx: active nav tab had no aria-current
- response-section.tsx: the status-code dropdown had no
  aria-haspopup/aria-expanded/role, and no Escape-to-close
- workflow-preview.tsx: same focus-trap gap as lightbox.tsx on the
  expanded-canvas modal

SEO:
- page.tsx: generateMetadata's hreflang/canonical URLs used a naive
  String.replace to strip the locale prefix, which also matched "/en"
  inside unrelated slugs (platform/enterprise, integrations/enrich,
  platform/self-hosting/environment-variables), corrupting those pages'
  canonical and alternate-language URLs. Replaced with a prefix-only strip.
- structured-data.tsx: the SoftwareApplication JSON-LD block compared
  url === baseUrl (no trailing slash) against the homepage's actual url
  (always has a trailing slash), so the condition was always false and
  this structured data never rendered anywhere, including the homepage.
- structured-data.tsx: "Mothership" in the indexed featureList violated
  the constitution's required language (the agent is "Sim", the surface
  is "Chat") - this ships in JSON-LD search engines parse.

The panel split unmounted AskAIPanel entirely on close, discarding
useChat's message state - reopening always started an empty
conversation, unlike the original single-component layout where
useChat lived in a component that never unmounted.

Fixed by keeping the panel mounted (via a hasOpened flag that never
resets) once first opened, and having the panel itself return null
when closed rather than being conditionally removed from the tree by
its parent - hooks still run every render, so useChat's state persists
across visibility toggles. The dynamic import still only fires on the
first open, so the initial-load win is unchanged.

Verified via a real click-through (open, type, close, reopen): input
persists correctly, and the panel chunk still has zero network
requests on initial page load. Performance unchanged at 75.

handleOpen set open=true synchronously, hiding the trigger button
before the dynamically imported panel had a chance to render anything
(next/dynamic renders null by default with no loading option) - on a
slow connection neither the button nor the panel was visible.

Added a loading fallback in the same fixed position so there's no gap
between the button disappearing and the real panel appearing.