regression: change in overflow checking

[Mark-Simulacrum]

The regression in diesel-softdelete is separate and regressed in #100757

Looks like we changed the boundary here which seems like it causes needless breakage for users - not convinced we should do that.

Originally posted by @compiler-errors in #103243 (comment)

[compiler-errors]

@ouz-a I think we might need to revert #100757 because of this issue and #104322 (and #104606 mentioning this too) -- do you want to do that, or should I?

[ouz-a]

@compiler-errors this is going to introduce small(compared to issues it raised) regressions, but I can revert it.