yes I also totally overlooked this. We have a check in the code that fires at the end of the SCF calculation, which was here never reached. In the future VASP version we will also have a check right after the delay phase (first 5 DAV steps) to make a first test of that. So this would be reported to the user much earlier in future.
That makes a lot of sense! I could have sworn that normally I should have spotted that issue, so I was surprised I did not here. Perhaps it would be good to have an option (accessible through an input boolean parameter) to have VASP stop when the check is triggered after the delay? Personally, I feel VASP could run considerably more robust as I easily keep finding new ways of breaking it, which often could potentially be avoided. Of course, I know it is also the user's responsibility. But I don't expect VASP to, e.g., have infinitely long lattice vectors when something went wrong with the POSCAR (if I recall correctly, it would read NaN "just fine", but this was a while back so maybe it got fixed).
Very good catch! Thank you I will add information on the wiki for that and crosslink to the TAUCAR page.
Small additional point: The TAUCAR page does mention that it is only from v6.6.0, the LTAU page does not. Unfortunately, I overlooked that point for a bit...
Again thanks for all the help, I hope that all of this will also help to resolve some nasty convergence issues for non-collinear mGGA calculations that I have been running. I guess overall the METAGGA page needs some updating after all the recent (very welcome!) changes and improvements. For example, it does not mention IVDW_NL either or what you mentioned earlier about that LMIXTAU might actually not be helpful (although I still do not understand when it is and when it is not).