This was fixed by !4980 (closed) and will be present in 9.0.2. I think the regression test T19345 covers the same ground, so we don't have to add another one.
It is very nice indeed that it is already fixed, but what about mitigation plan and public communication?
It looks like any library dealing with fromIntegral and Natural is affected, silently returning an incorrect result. This is outright dangerous. And almost invisible in testing, because QuickCheck tests rarely generate inputs of such magnitude.
In my case the bug hit two libraries, both with cryptographic applications, and I was lucky to have thorough enough tests. What is the best way to fix them?
Simplest might be to refuse to comipiple with __GLASGOW_HASKELL == 901, users will have to use 8.x or >= 9.0.2. I am not sure that more complex work-arounds are warranted.
Yes, one way around is just skip GHC 9.0.1, probably by base /= 4.15.0.0 in cabal file. I wonder if there are other, less disruptive options, but they require a way to audit external dependencies. Potentially one can run a build logging triggered rewrite rules, and compare logs between GHC 9.0.1 and GHC HEAD. It does not sound quite appealing, but once affected libraries are identified, they can be fixed with
Orthogonally to 9.0.2, I believe it is worth to warn public about the issue, before more people get hurt. IMO it would be more appropriate if such warning comes from GHC developers, even if in form of email to Haskell Cafe or whatever.
It might very well be that my perception is biased; if others do not share my feelings, feel free to close, I do not mean to dictate others what to do.