The border is not a test
An accent colour earns its place when it can carry mass — when it holds a button face, a notification badge, a selected-state background — and still reads as itself rather than as noise. Specifying it only on two-pixel borders is not using a colour; it is deferring the decision. The border does structural work (it separates, it contains) but almost none of the chromatic work that justifies introducing a third hue into a system built on two.
The problem surfaces predictably at scale. Something in the interface graduates from a thin rule to a filled chip — a tag, a pill, a progress bar — and the accent, never tested at area, either screams or dissolves. If it screams, someone mutes it with opacity, which produces a muddy near-neutral that reads as a mistake, not a tone. If it dissolves, someone darkens it until it no longer matches its own two-pixel self. The colour has now forked: one value for lines, another for fills. That fork is a symptom, not a solution.
There is a simpler discipline: before a colour joins the palette, put it on a rectangle roughly the size of a button. Then on a rectangle the size of a card header. If it cannot sit comfortably at both scales — if it needs opacity correction, or a shifted lightness value, or a darkened border to stabilise it — it is not ready. It is a rule masquerading as a palette entry.
Two colours far enough apart in hue are still not a palette if one of them cannot travel. Distance is necessary; range is too. An accent that only ever appears thin is a tic, an aesthetic signature borrowed from somewhere, applied without asking whether the colour beneath the affectation can actually work. Borders are useful. They are just not evidence.