Performance

How Nearer stays fast on a phone: less to download, fewer reads before the first screen, and small caches that clear quickly.

What loads first

The hero photo has loading priority. Other images do not.
The window and the courtyard read in two database waves. The build log notes this was six and eight waves before.
The catch-up sentence is bounded, so a model call never holds the first screen.
Past runs in the build log record courtyard LCP 6.0 s → 1.2 s with CLS 0.37 → 0, and 18 → 6 API calls. Window LCP 3.6 s → 0.7 s with CLS 0. Images 312 → 82 KB in the same pass.

Pages are prerendered and served from static assets. Nearer uses no revalidation, so no extra cache bucket is needed.
Link previews use static images. The recipe page title and description are still per recipe.
Ritual configs are bundled at build time.
The few bundled demo photos are served as-is from static assets, edge-cached. There is no built-in image optimizer in this setup.
Reliability helpers noted in the build log: upload queue with retry, stale-while-revalidate, per-section boundaries, and a service worker.

Reads come from the database first on every public view. On top of that:

Cache	Where	Rule
List cache, `useJson`	Browser memory and `sessionStorage` (`aangan:json:*`, small payloads only)	Any successful write drops all of it; lists refetch once 200 ms later, coalesced. A 401 drops it. Sign-in, sign-out, or switching family or person drops it. Between a write and the refetch is about a quarter of a second.
Cards snapshot, courtyard payload	Browser memory	Same write and auth-change invalidation as above.
Last first-screen photo, `aangan:hero`	Browser	Set when the window loads. Removed on auth change, so the next family never paints the last family's photo.
Weather, geocoding, holiday feed	Server, shared across isolates	Cached with deadlines. The weather cache notes 30 min / 24 h / 10 min for a miss. Public data only.
Model output cache	Server	Cached outputs may be recomputed per isolate.

Shared JavaScript means what every route loads: the build's root main files plus the chunks of the root layout. It is measured gzipped, as a server would send it. The default budget is 220 KB. The build log records shared JS 272 → 202 KB gz.
Images use responsive sizes, AVIF/WebP, and blur placeholders, with no layout shift.
Fonts load with subset and display: swap. Two families only.
Motion loads through LazyMotion with domAnimation, so the motion bundle stays small.
Waveforms are plain DOM bars, 64 or fewer, with precomputed peaks. The live analyser runs only while recording or playing.
At most two backdrop-filter layers are visible at once, with a solid fallback when unsupported.

node scripts/perf-budget.mjs checks shared client JS against the budget. It fails when the total is over budget, and when the manifests do not say where the shared chunks are.
scripts/perf-vitals.py checks page weight and Core Web Vitals.
Targets in the design system: Lighthouse mobile performance ≥ 90, accessibility ≥ 95, interaction under 100 ms.
scripts/walk.py fails on sideways overflow, text over 34px outside data-display, and text under 12px at 360 and 390pt.

Written by Muse Spark (Meta) from the build log on 2026-11-02.