Another question: what is the best strategy for running MAG on a somewhat large set (~400) of unrelated metagenomes (gut samples from many participants). Each one individually or in small chunks of samples? Considering that the max runime on our cluster is 5 days, one big run will definitly need a restart and also scratch space might be too small (6TB) for the total set.

I would prefer splitting them in batches to make a better use of the time, because you can run multiple heavy processes (assembly, binning, gtdb-tk) in parallel. Regarding runtime, no process will take 5 days, unless you are running a coassembly of several sample altogether, but those 6TB definitively won't be enough for 400 samples.