It's proving very difficult to get rid of glitches, so here are some more ideas:

Median filter the whole timestream (downsampled) at a resolution that will pull down the glitch peak. Subtract out the median-filtered timestream from the original, and look for outliers in that distribution: in principle, those should be glitches.

Another option: subtract out the noise map AND the best-model map from the original timestream before trying to pull out additional astro model stuff. The map-to-timestream astro model can't include glitches because they're averaged over, but the sky-subtracted timestream DOES include glitches (no matter how many PCA components are removed) because glitches are NOT correlated across detectors. The weird thing is that this is effectively subtracting out exactly what was calculated from the sky subtraction - the question is whether subtracting it out BEFORE sky subtracting again gets you any benefit. If it was just noise, subtracting noise from noise is fine, but the "noise" does include some residual signal.