The ellipse() grammar defined in Shapes was clearly designed to resemble the grammar of radial-gradient(), but it's different in small but important ways, for no apparent reason.

In particular, the ellipse part of the radial-gradient() grammar is: <extent-keyword> | <length-percentage>{2}. You either say something like "farthest-side", which defines an ellipse, or give lenpers for both axises.

On the other hand, the ellipse() function says [ <extent-keyword> | <length-percentage> ]{2}? - for each axis, you can either provide a keyword or a lenper. So you can do something like ellipse(farthest-side closest-side), to make its width the amount necessary to hit the farthest horizontal side, and its height the amount necessary to hit the closest vertical side. But you can't do ellipse(closest-side).

Given the seniority of the radial-gradient() syntax, and the lack of major issues with it, I think we should converge on that as the grammar for ellipse().