Since you are integrating a rational function with an endpoint singularity, which quadgk cannot integrate efficiently since that point is not smooth, I believe you are seeing how the adaptive grid constructed by quadgk resolves the distribution of floating-point numbers, which are clustered logarithmically about the origin. If you reverse the interval, as you did, or just shift it like this
julia> quadgk(s->(s-1)^(-0.9),1,2)
(9.75587242705648, 0.011981660960530958)
the singularity is located at a point where the Float64s are less dense and quadgk will give up as soon as it has subdivided intervals so close to the singularity that the quadrature nodes it want to use are rounded off to the endpoints. Since evaluating at the endpoint, which is singular, would give infinite error quadgk returns early and with an error estimate greater than your tolerance.
If you want to fix this, consider a change of variables to cancel the endpoint singularity or a specialized quadrature scheme for kernels with endpoint singularities in mind. Most likely they will be more accurate and more efficient, although it may take some work to find the best strategy for your kernel. One possibility for kernels with power-law singularities is to use Gauss rules for Jacobi weight functions, which you can compute with QuadGK.jl. I’ve also tried to automate h-adaptive Gauss-Kronrod rules for Jacobi weight functions here with some examples here, but the Kronrod rules may or may not exist for all power law divergences.