Paint the marks in a colour that can be seen

The unification shipped three washes calibrated against the wrong number. The
old rule's visible part was an outline and a 4px bar; its background was
accent at 16%, which was doing nothing, and that is the value the new
percentages were reasoned from. Composited over the page, decided prose
rendered #f2f9ff and the pointer #daeeff — white, and nearly white. Nothing
failed, so nothing was logged; the ranges registered correctly and were painted
in the colour of the paper.

Ring rather than accent for the blues: accent is oklch(0.85 0.09 250), already
pale, so no percentage of it reads. Ring is 0.62 0.14 258. The pointer now
lands at #b7cff2, about the weight of the browser's own text selection, which
is the right strength for something standing in for an outline. A live comment
sits within a shade of the amber that was demonstrably visible before, and a
decision keeps the pointer's hue so touching its card deepens one colour rather
than swapping it for another.

background-color rather than background: only a few longhands are permitted on
a highlight pseudo-element, and the shorthand works today solely because
browsers expand it before filtering.

The rendered hex is recorded beside each declaration, and a test now does the
arithmetic: it resolves each tone against the theme, composites it over the
page, and refuses anything too close to the background to be a mark. There is
no DOM and no CSS.highlights in the test runtime, so a mark painted in white
is indistinguishable from one that failed to paint — which is how this shipped
twice. It also checks the recorded colours, because a stale comment is what the
next person would calibrate against.