My best practice would be to do the 12 individual samples. After all, you have two major sources of variation: a) the library preps of each of the 12, and the conditions of each of these. Both you do typically not want reflected in clustering and visualization. You want celltypes cluster by celltypes. DE analysis, using the raw data will then take care of the condition differences. That's at least how I have been doing this over the years. I don't think overcorrection is really.a thing here. After all, you probably will have a good idea which celltypes to expect, so check whether these can be found and separated. You could start by assigning celltypes using reference labels, e.g. using SingleR first, so this whole clustering-UMAP typical workflow is deemphasized and you immediately see which celltypes are present.