User avatar — accessibility

Keyboard

Not focusable and not interactive. In the account menu or a picker the host button is the control.

The face is aria-hidden="true" and carries title with the user's name. It is a compressed, decorative rendering of a name that is already in the DOM beside it.
Consequence: like every mark in this system, it must be accompanied by the name in text. A face on its own is silent to assistive technology, and unlike initials it does not even encode letters.
An uploaded photo is an <img alt=""> for the same reason.

Nothing focusable, so no focus ring. The surrounding control keeps the foundation ring.

Computed from the palette tokens:

The eyes and mouth use the paired --uav-N-ink for each colour, so the drawing never sits on a ground it cannot be read against — the two light colours (yellow, orange) get dark ink, the three dark ones get white.
The face is decoration: nothing in it is load-bearing text, so it is not held to a text contrast ratio.

The distinction is visual only. "Colleague" versus "contact from outside" is carried by the drawing; a screen-reader user gets the name and whatever the surrounding text says about the role. Where the difference matters for a decision — attendee lists, owner fields — the text must say it, not the face alone.
Two colleagues can land on a similar colour pair; the face is a memory aid, never an identifier.
The drawing is generated, so it cannot be described. alt/title carry the name, and that is all a reader gets — which is the same amount of information an initials avatar gives.