Ruling out the obvious suspects first

CrUX lag? No — that affects Search Console's field data, not a live local trace in DevTools. Different data source entirely.

Capture method? Confirmed I was already using the Performance panel's dedicated reload-and-profile button, not a manual refresh — same requirement as the CLS work, where the purple shift marker only shows up reliably under a proper cold-start trace. Ruled out.

Lighthouse cross-check? This is where it got interesting. Lighthouse didn't just fail to find an LCP — it failed outright:

Clearing the browser cache timed out. The page did not paint any content. (NO_FCP)

No First Contentful Paint at all means something blocked rendering completely under Lighthouse's test conditions — a much bigger red flag than a missing metric.

The Cloudflare detour

The site runs a Cloudflare rule scoped to US traffic only. Google's testing infrastructure doesn't reliably originate from a US IP, so the working theory became: Lighthouse is hitting a challenge page, not the real site. That checked out — the rule fires as a challenge, not a hard block, which explained the Lighthouse failure well enough to stop chasing it.

Two real, unrelated problems, stacked on top of each otherCrUX lag and the Cloudflare geo-rule were both legitimate issues worth knowing about — but neither one was the cause of the blank local trace. Easy to conflate two real findings into one wrong conclusion. Worth separating explicitly before drawing any.

With Lighthouse and PSI both ruled out as unreliable for this specific test, local DevTools — running as a real US-origin session — became the only trustworthy read.

Back to the console

Opened DevTools Console on a clean reload and just read what was actually there:

tweetie.js:1  Failed to load resource: 404 application.js:1161 Uncaught TypeError: gfx(...).twittie is not a function js?v=3.exp&sensor=false&.js:204 Uncaught InvalidValueError:  Map: Expected mapDiv of type HTMLElement but was passed null. jquery.carouFredSel-6.2.1-packed.js:15 carouFredSel: Set a width for the items! jquery.carouFredSel-6.2.1-packed.js:15 carouFredSel: No element found for ".testimonials".

The Google Maps error was already a known issue — a legacy, key-less Maps script throwing errors on pages with no map element. Already on the fix list.

The tweetie.js 404 and its cascading TypeError were new, and looked promising: an uncaught error mid-script could plausibly halt anything scheduled after it in the same synchronous block, including carousel init.

But the line that mattered most was the easiest one to skim past:

carouFredSel: Set a width for the items!

Not red text. Not an error. Just a warning — the kind you scroll by a hundred times while hunting for the "real" problem.

Why the warning mattered

Every carousel on the site used the same CSS, including the ones already fixed for CLS. That earlier fix had locked down height, overflow: hidden, and white-space: nowrap — solving the vertical half of the layout-shift problem. But nothing in that CSS gave the <li> items an explicit width. They were display: inline-block with no defined width, shrink-wrapping to content until JavaScript resolved actual dimensions.

CarouFredSel resolves that ambiguity itself, by measuring rendered widths at runtime and writing them back in as inline styles:

<li style="width: 274px;">

Inspecting the live carousel confirmed it: every item had a JS-written inline width, and the wrapping <ul> had ballooned to a computed width in the thousands of pixels — the sum of every item at its shrink-wrapped size, before JS narrowed things down to the real carousel width.

That's a second layout pass happening after initial paint. And critically, it wasn't showing up as CLS. The working theory: because the container height was already locked by the earlier fix, and everything sat inside overflow: hidden, the resulting shift was purely horizontal and largely off-screen — invisible to CLS's scoring. LCP, on the other hand, kept moving between the pre-JS shrink-wrapped state and the post-JS resolved-width state, and DevTools couldn't cleanly attribute a stable LCP candidate to something still being recalculated.

The fix

One line:

ul.latest-work.disable-select li {  display: inline-block;  white-space: normal;  height: 155px;  width: 274px;  vertical-align: top; }

Hardcoding the exact value carouFredSel was already calculating and injecting via JS — just moving it from a post-load patch to a pre-load CSS declaration, so the browser has everything it needs in a single layout pass.

Result: the <ul> stopped inflating to absurd computed widths, and the previously blank LCP reading came back clean on the very next trace.

The actual lesson

The CLS fix and the LCP fix looked like the same category of problem — "carousel causes layout instability" — and it would've been easy to assume one fix covered both. It didn't, because they were sensitive to different axes of the same underlying issue: vertical instability vs. horizontal instability, and only one of Google's two metrics was built to notice the one that had been left unaddressed.

Two things worth remembering from this one:

  • A console warning isn't automatically low-priority just because it isn't red. Set a width for the items! sat in every console log on every carousel page, ignorable-looking, for as long as the carousels existed. It was the whole answer.
  • Rule out your testing environment before you rule out your code. The Cloudflare geo-rule genuinely was interfering with Lighthouse — real, and worth fixing on its own — but it had nothing to do with the LCP problem. Confirming that early kept the investigation from chasing infrastructure when the bug was sitting in a <style> block the whole time.

Next up was rolling the same width fix across the rest of the carousel partials sitewide, then a full reload-and-profile sweep to confirm it held everywhere — followed by giving Search Console's 28-day CrUX window time to catch up now that there was finally good data for it to average in. (Update: it has — mobile and desktop are both clean now across the board.)