Cred stuffing can often be done with 5 or less "known" passwords used by the victim. When the attackers have enough IPs (via proxies or cloud providers - remember that most cloud providers give out multiple ipv6 /64s no questions asked), It'll end up looking like any other login attempt from a consumer VPN where the user forgot their password and is trying 3 or 4 different ones.

Not that easily. A sophisticated attacker is going to be using a lot of different throwaway IP addresses. One of the common ways to (try to) prevent this are often how you end up with stuff like "all IPs we think are AWS or data centers or [...] are blocked from our service."

How many logins/second from how many different IP addresses and browsers would you have to try in order to achieve this in a way that is hard to detect? How hard is it to rent a botnet that would mean IP-based filtering would have too large a blast radius?

It also assumes malicious request have same or similar signature attributes (ex: IP addr, http hdrs), which is not always the case.

They probably could. Assuming that these attempts aren’t executed using a botnet with dozens of thousands of machines.

This assumes that people have good monitoring over the login pages/endpoints.