Dark dashboard parity requires page-level composition, not token swaps

Observed 2026-09-08 while implementing COR-4091 in ender-inference from a user-supplied visual reference and synthetic localhost CUA checks. Replacing green-tinted tokens was insufficient: the dashboard still read as sparse cards floating on black until each active route received one dominant warm-charcoal parent surface, with nested cards only one neutral level apart. A connected desktop stepper, vertical mobile stepper, and an intermediate-width filter reflow removed the remaining composition and overflow gaps. Retaining every baseline selector group and checking computed page/canvas colors plus 390px scroll widths provided a practical guard against behavior regressions while changing only CSS. Local evidence: commit 699bcfd2b35493c940ff13521082369187c2f9df; no staging or production observation.