Vaguely remember BOINC itself would not run more than N cores out of the box, 48 sticks in my mind. Anyway, not so long ago learned that BOINC does not have more slot handles than 400, so there's your per-client cap. Since with those counts it's a come and go of jobs, some left empty after completing a job, the practical limit is lower, jobs completed, ready to upload, new jobs started in meantime. There were reports of slots not being fully emptied, and thus BOINC not putting them back into the availability list and eventually running out of slots to start new jobs with "[error] exceeded limit of 400 slot directories". A different magnitude problem of footing the associated electricity-bill.

Cool. The theory was, that after a certain client version, the N CPUs would be ignored and there would only be listening for the % of processors setting. As you were "it seems to have solved itself", no ncpus setting in the cc_config.xml, guess something got beaten into submission to the % of processors rule, and it purely was the initial 'getting started' cap (thought it was 5 per device, but no less than 1 per core). The obscure sides of BOINC... be as cryptic as possible to those that do find the event log event-ually ;o)

BTW, the techs in the control room have in past set exceptions for individual device [on request]... yes, they've got handles down to the single host. Largest ever seen for WCG at BOINCstats was I think 96 cores., but now there's one with 127 (an odd number... pun): https://boincstats.com/en/stats/15/host/list/ . There's multiple with 63.5 (127) in hyperthreading.