Web Vitals

The five metrics collected, how they are measured, and where they surface.

On by default. No library, no polyfill: the tracker reads PerformanceObserver directly, which is why the whole feature costs a few hundred bytes.

Metric	What it is	Unit
`lcp`	Largest Contentful Paint	ms
`cls`	Cumulative Layout Shift	unitless
`inp`	Interaction to Next Paint	ms
`fcp`	First Contentful Paint	ms
`ttfb`	Time to First Byte	ms

When they are sent

Once per document, on the way out: the first time the page is hidden or unloaded. That is the only honest moment: LCP is not final until the page stops loading, and INP is not final until the reader stops interacting.

They travel as a web_vitals event with the values under e:

json

{ "s": 1, "t": "web_vitals", "u": "https://example.com/docs/x", "e": { "lcp": 1180, "cls": 0.04, "inp": 96 } }

A metric the browser did not report is simply absent. Nothing is estimated and nothing is filled in with a zero, so a p75 is computed over the pages that actually measured it.

How each is measured

LCP is the last largest-contentful-paint entry, using renderTime where the browser gives one.
CLS is the worst 5-second session window, the way the metric is defined: shifts within a second of each other and inside five seconds of the window’s start accumulate, and a shift after user input is ignored.
INP is the longest interaction over 40 ms, read from the event timing entries.
FCP and TTFB need no observer at all: both are already on the navigation timeline when the page hides.

Browsers that do not implement an entry type report nothing for it. Safari has no INP today, so an INP p75 is a Chromium-and-Firefox figure, and the report says how many samples it is built on.

Turning it off

html

<script defer data-site="1" data-vitals="false" src="/mf.js"></script>

Where they surface

GET /api/stats/performance returns p50, p75 and p95 for each metric, and it takes the standard query envelope, so every one of them can be cut by any dimension: by page, by country, by device type, by browser.

That last part is the useful bit. A site-wide LCP is a number to feel bad about; LCP on your five most-read pages, on mobile, is a number you can act on.

What it does not do

There is no synthetic run, no lab measurement and no scheduled audit. These are field metrics from real readers only. A page nobody visited has no Vitals, and Micaforge will show it as absent rather than as zero.