Typefaces the claimed OS ships as separate faces measure as separate faces

A font request is answered by the host's font stack, not by the browser and not by anything the user agent sets, so changing navigator.platform does not move a single glyph. That makes the font layer an OS oracle — but the obvious way to read it does not work, and the reason is worth stating because it is what shapes this check.

Asking whether a named Windows face such as Segoe UI or Calibri is INSTALLED returns yes on a Linux machine that has none of them: fontconfig ships metric-compatible substitutions, mapping Calibri to Carlito and Cambria to Caladea, and those clones are built to match the original's advance widths exactly.

Measured on stock Chromium on Linux, Calibri and Carlito both render a probe string at 853.559px, and Segoe UI returns 914.273px — the identical value a real Windows Chrome returns. A substitute designed to match the original's metrics cannot be detected by measuring the original's metrics, so no width table and no presence test can see it. What a substitution scheme cannot do is provide a distinct clone for every face.

Faces without one fall back to the platform default and therefore collide WITH EACH OTHER: on that same Linux host, Tahoma and Arial both measure 905.203px, because Tahoma has no substitute and lands on the default sans-serif that Arial already maps to. On Windows those are two different typefaces and cannot share an advance width.

So this row asks a self-referential question — do faces the claimed operating system ships as SEPARATE typefaces actually measure separate here — and it carries no expected widths at all, which is what keeps it from rotting when a font version changes. Two refusals are built in and both were measured rather than reasoned about.

First, collapse onto a generic is never scored: on Windows Chrome the default monospace IS Consolas and the default sans IS Arial, so `Consolas === monospace` and `Arial === sans-serif` are true on a genuine Windows machine, and a check reading that as a missing font would fail every real Windows browser. Only collisions between two non-default named faces count.

Second, a browser that defeats font enumeration flattens every metric, and pervasive collision is that defence rather than an OS lie — so the row requires a majority of the probed set to be mutually distinct before it will assert, and declines on a flattened palette. A face that does not exist is probed as a control; if it resolves to its own metrics the browser is answering synthetically and the whole reading is discarded.