I wasn't able to conclude if [Andy broke something] or if our use of Base.hash_integer assumes too much

perhaps a little of both. I am not seeing that it is, pedantically speaking, a documented requirement that hash_integer coincides with hash. but this is probably just due to the lack of documentation rather than an intentional freedom for them to diverge.

in this case, I think changing

function hash_integer(n::Integer, h::UInt)
    h ⊻= hash_uint((n % UInt) ⊻ h)

that first line to h = hash((n % UInt), h) will recover the coincidence. I can submit a pr shortly

FYI (maybe better as a separate issue) if you're interested, there may be potential to use the new hash_bytes to hash really huge numbers for a big performance win, if that's ever a bottleneck for you. I once worked on a project where each number was Gb large where such an improvement might have been relevant.

in this case, I think changing

function hash_integer(n::Integer, h::UInt)
    h ⊻= hash_uint((n % UInt) ⊻ h)

that first line to h = hash((n % UInt), h) will recover the coincidence.

seems to work locally. I also have a local fix that adapts the hash_integer special case for BigInt accordingly. I'll open a PR with both changes in a few

Triage decided that this is not an issue as hash_integer is internal. See #58387 (comment) for the triage comment