Steffen reminded me over email that the particular machine only has a 1 dalton accuracy, and that the 150ppm parameter setting is somewhat inappropriate. As seen yesterday, it works fine for larger peaks, but fails for low intensity peaks. So, I reran the centWave peak detection with 750, 1000 and 1250 ppm, and that indeed make XCMS recover many more metabolites, and, also important, with more extracted ion chromatograms per metabolite, yielding a more accurate mass spectrum. At the same time, I notice that profiles are not as clean as before, but that's where the peak fitting with (Modified) Gaussians come into play.
The original 150ppm results:
The 750ppm results:
And for 1000ppm (1250ppm did not further improve):