[graphitemaster]

Author here. I believe many of the complaints here are concerned about what is more common rather than what is universally better over all possible inputs which is actually the point of view this article is written from. The issue with the "general case" is that exploits are never actually found there, they're always found in the edge cases such as large, or "pathological" values as it were. Signed integer arithmetic has more of these edge cases than unsigned when it comes to sizes, indices, and offsets used in expressions controlling memory (either directly, or indirectly) which is the most common application of integers in a codebase. The native-word-size unsigned integer type covers the full numeric range for those operations, while signed simply cannot. At the same time, preventing a whole numeric-range that is simply incompatible with those operations (negative values), and a whole class of issues related to undefined operations. The only real edge-case unsigned has that is more error-prone is values close to zero under subtraction and that's relatively easy to account for which I go into great detail to explain.

[mpweiher]

Which is why the only sensible option is a proper numeric tower.

Then the common cases are fast and the edge cases slightly slower, but everything works the same.