EDIT:
I'm aware I could just add a compensating "$-c_1(0)$" term to get rid of the vertical offset, but that feels not in the spirit of Fourier series... and doesn't explain the mystery anyway.
Into Desmos, I plotted:
$$f(x)=x^k\\c_1(n)=\frac{1}{\pi}\int_{-\pi}^\pi f(x)\cos(nx)\,dx\\c_2(n)=\frac{1}{\pi}\int_{-\pi}^\pi f(x)\sin(nx)\,dx\\\mathcal{F}(x)=\sum_{n=0}^Nc_1(n)\cdot\cos(nx)+c_2(n)\cdot\sin(nx)$$
The $c_1$ and $c_2$ functions are coefficient-finding functions, and rely on Desmos' numerical integration.
And this Fourier series $\mathcal{F}(x)$ fits $f(x)$ very very well, with more accuracy as $N$ increases, only when $k$ is odd. When $k$ is an even integer, the series fits the curve very well still... but at a large y-offset. The series oscillates around some y-intercept, whose value as a function of $\pi$ I have not been able to experimentally determine. For reference, here are two images, the first of $k$ odd, the second of $k$ even. The green lines are the original function, and the red lines are the Fourier series.
$k=3$ and the Fourier series fits well:
$k=4$ and the series fails... bizarre vertical offset.

Many thanks for any suggestions about where this comes from, and how to correct it.