I think in that case it might a matter of length, rather than symbols, but I've never put in the effort to pin down the problem exactly, now that my password manager settings work.
Like, I caught them doing the "ask for FQDN but reject FQDNs" a couple of weeks ago in either Route53 or something immediately adjacent (i.e. where people who know better should have been involved), yesterday SageMaker stopped launching kernels because of a "known issue" (don't worry, it doesn't meet the definition of downtime though), and today I had to diagnose an unresponsive control panel by opening the console and digging up a REST error response that they didn't bother to handle in any way.
Everything about AWS is not merely broken and half-assed, but fractally so, so that even after using a service for years you are still uncovering severe caveats and awkward limitations. Password validation traps fit that pattern to a T.