Web Vitals Budgets — Media-Heavy Surfaces

wp-performance explicitly opts out of browser measurement. This skill is the browser half: real LCP/CLS numbers via Chrome DevTools MCP or Playwright, with budgets for the media patterns this site actually ships.

When to use

• An above-the-fold media change (hero video, animated header, image strip, lockup image) is about to ship or just shipped to skyyrose.co — measure before AND after.
• A launch audit needs LCP/CLS/INP numbers with element attribution.
• A CSS/JS change altered above-fold layout — CLS must be re-measured, not just visually QC'd.

When NOT to use:

• Backend slowness (TTFB, REST, admin, cron) → wp-performance. This skill only attributes TTFB inside the LCP breakdown; it does not fix it.
• Changes with no above-fold rendering impact (PHP-only logic, admin screens, email templates).
• Judging design quality — this skill produces milliseconds and shift scores, not taste.

Inputs

Absent input = STOP. Numbers, not vibes — no browser tooling means no verdict.

1. A reachable target URL (production https://skyyrose.co/... or a preview). Probe it first (see Verification check 1); a non-200 or sub-50KB response means measure nothing yet.
2. Chrome DevTools MCP or Playwright MCP available. If neither is available this session, the LCP/CLS/INP checks are SKIPs, not PASSes — say so explicitly and name who closes them (the caller re-runs with browser tooling). Do not substitute a curl for a CLS number.
3. The budgets table below — a measurement without a budget to compare against is trivia.
4. Logged-OUT context — the WP admin-bar shifts everything 32px and corrupts CLS.

Budgets (skyyrose.co, mobile 4G reference)

Metric	Budget	Hard fail
LCP	≤ 2.5s (excl. known TTFB debt ~2.3-3.4s — track separately, don't hide it)	> 4.0s
CLS	≤ 0.05	> 0.1
INP	≤ 200ms	> 500ms
Above-fold media weight	≤ 1.5MB total	> 2.5MB
Header animated asset	≤ 400KB	> 1MB

Procedure

1. Baseline TTFB + reachability with cache-busted curl (Verification check 1). Record whether the edge cache HIT or MISS — the server-timing/x-ac headers say which; a MISS number and a HIT number are different experiments and must not be compared.
2. Chrome DevTools MCP: performance_start_trace → navigate cold (cache disabled) → performance_stop_trace → performance_analyze_insight for the LCP breakdown (TTFB / load delay / load time / render delay). Mobile viewport 390×844 AND desktop 1440×900.
3. Playwright fallback: browser_evaluate → new PerformanceObserver for largest-contentful-paint + layout-shift entries; report element attribution for LCP (WHICH element — if the header video steals LCP candidacy from the hero, that's a regression even at the same ms).
4. Three runs, report median. One run = noise.
5. Before/after on the SAME network conditions — deploy comparisons across different cache states are invalid.
6. Report in the format: LCP 2.31s (el: img.hero-lockup, ttfb 1.9s) | CLS 0.02 | INP 140ms | mobile, 3-run median, cold — element attribution + conditions or the number doesn't count.

Verification

A trace that errored or timed out is an artifact, not a green result — re-run it (bug-230). Each check states its pass condition; each can fail.

curl -s -o /dev/null -w 'code=%{http_code} ttfb=%{time_starttransfer}s total=%{time_total}s size=%{size_download}B\n' \
  "https://skyyrose.co/?cb=$(date +%s)"

PASS: code=200, size ≥ 50000B. Record ttfb — it is the floor under LCP (LCP can never beat TTFB + render). Observed 2026-07-28: code=200 ttfb=1.888431s total=2.092811s size=136605B on an edge-cache MISS. [live]

// Playwright browser_evaluate — run on the target page, logged out, mobile viewport
new Promise((resolve) => {
  let lcp = null, cls = 0;
  new PerformanceObserver((l) => { const e = l.getEntries().at(-1);
    lcp = { t: e.startTime, el: e.element?.tagName + '.' + (e.element?.className || '') };
  }).observe({ type: 'largest-contentful-paint', buffered: true });
  new PerformanceObserver((l) => { for (const e of l.getEntries())
    if (!e.hadRecentInput) cls += e.value;
  }).observe({ type: 'layout-shift', buffered: true });
  setTimeout(() => resolve({ lcp, cls: +cls.toFixed(4) }), 6000);
});

PASS: median of 3 runs → LCP ≤ 2.5s AND the attributed element is the intended hero (not the nav video), AND CLS ≤ 0.05. Hard fail at LCP > 4.0s or CLS > 0.1 per the budgets table. [live]

curl -s -o /dev/null -w '%{size_download}\n' "https://skyyrose.co/<header-asset-path>?cb=$(date +%s)"

PASS: header animated asset ≤ 409600 bytes (400KB budget); above-fold media sum ≤ 1572864 bytes. [live]

Prove the CLS check can fail (rule 3): once per harness setup, strip width/height off the hero <img> in a local copy and re-measure — CLS must go red — then restore. A CLS gate never observed failing is a guess with a citation.

Worked example

Real probe of production, 2026-07-28, from this repo:

curl -s -o /dev/null -w 'code=%{http_code} ttfb=%{time_starttransfer}s total=%{time_total}s size=%{size_download}B\n' \
  "https://skyyrose.co/?cb=$(date +%s)"
curl -sI "https://skyyrose.co/?cb=$(date +%s)" | grep -iE "^(HTTP|x-ac|server-timing)"

Observed output [live]:

code=200 ttfb=1.888431s total=2.092811s size=136605B
HTTP/2 200
x-ac: 1.sjc _atomic_bur MISS
server-timing: a8c-cdn, dc;desc=sjc, cache;desc=MISS;dur=2493.0

Reading it: the cache-busted request MISSed the WP.com edge (cache;desc=MISS;dur=2493.0), so 1.888s TTFB is the origin cost — consistent with the known ~2.3-3.4s TTFB debt band. Any LCP measured on this cold path carries that debt; report it separately (ttfb 1.9s inside the LCP line) instead of hiding it, and also measure the warm-cache path real visitors mostly hit. The browser half (LCP element attribution, CLS) still requires the Playwright/DevTools run — this curl alone is [live] evidence for TTFB and reachability ONLY, never for LCP/CLS (bug-287: evidence scope must cover claim scope).

Failure modes

• CLS traps specific to this theme:
  • .min rebuild changing rendered sizes: any CSS edit that alters above-fold layout re-measures CLS, not just visual QA (production serves .min — measure after npm run build).
  • WP admin-bar (logged-in) shifts everything 32px — measure logged-OUT.
  • Batcache/edge cache serves stale HTML referencing new CSS versions (or vice versa) during deploy windows — measure after cache settles or with cache-busted URL, and note which.
• WebFetch on live HTML is banned — it strips <script>, so anything script-injected (LCP candidates included) silently vanishes. Use curl with ?cb=$(date +%s).
• Blank Playwright screenshots: call img.decode() on hero images before screenshotting, or the capture races the decode and you QC an empty frame.
• Header video stealing LCP candidacy from the hero at "same ms" — always check the attributed element, not just the number. Small + fetchpriority="low" on nav media; fetchpriority="high" on the ONE intended LCP element.
• Poster is the LCP candidate for hero video — optimize it like a hero image (≤200KB), preload="metadata", autoplay muted loop playsinline, explicit width/height.
• Auto-scrolling strips: content-visibility: auto off-screen, transform-only animation (compositor thread), loading="lazy" beyond first viewport, ≤100KB/image.
• A dead trace read as green (bug-230): DevTools trace that errors/timeouts has no findings because it has no data — re-run, never report it as passing.
• Scope jump (bug-287): a repo-side asset diff is not a live vitals claim — "regressed LCP" requires the [live] probe.