This paper might be helpful

Article Performing elemental microanalysis with high accuracy and hi...

it highlights the implortance of understanding error budgets in the analysis

"The relative errors (after normalization to 100 %) and the raw analytical total obtained with the k-ratio protocol and DTSA-II are shown at various locations. While accurate analytical results can be obtained at certain locations in image, at most locations in the images, the deviations from the correct stoichiometry are severe"

Thanks, Peter. That is a really interesting paper from the lead developers on DTSA-II at NIST.

Your quote is from a section warning the reader when EDS quantification is inappropriate, specifically with severe topology.

I think the abstract summarizes the thesis of the paper really well, which is that very good accuracy can be achieved with EDS despite its reputation of being semi-quantitative.

" Electron-excited X-ray microanalysis performed in the scanning electron microscope with energydispersive X-ray spectrometry (EDS) is a core technique for characterization of the microstructure of materials. The recent advances in EDS performance with the silicon drift detector (SDD) enable accuracy and precision equivalent to that of the high spectral resolution wavelength-dispersive spectrometer employed on the electron probe microanalyzer platform. SDD-EDS throughput, resolution, and stability provide practical operating conditions for measurement of high-count spectra that form the basis for peak fitting procedures that recover the characteristic peak intensities even for elemental combination where severe peak overlaps occur, such PbS, MoS2, BaTiO3, SrWO4, and WSi2. Accurate analyses are also demonstrated for interferences involving large concentration ratios: a major constituent on a minor constituent (Ba at 0.4299 mass fraction on Ti at 0.0180) and a major constituent on a trace constituent (Ba at 0.2194 on Ce at 0.00407; Si at 0.1145 on Ta at 0.0041). Accurate analyses of low atomic number elements, C, N, O, and F, are demonstrated. Measurement of trace constituents with limits of detection below 0.001 mass fraction (1000 ppm) is possible within a practical measurement time of 500 s. "

Regardless, that level of accuracy is overkill for my purposes. I simply need to compare a spectrum to a database of spectra and get the best match. I don't necessarily care about the stoichiometry .