We did not modify the bcrypt_rounds parameter. You mean the password hashing in bcrypt could drive a cpu% spike which the traffic management erroneously interprets as too high load?
Does that really sound realistic that the CPU is that busy with a single login and the following hardware - one VM for nodeBB and the same for mongoDB Debian GNU/Linux 13 (trixie), Kernel 6.12.57 4 Cores x86_64 Architektur 8 GB RAM