Annotate 3.14+ ipaddress version / max_prefixlen as Literal

[ABruihler]

On 3.14+, _BaseV4 / _BaseV6 declare version and max_prefixlen as bare Finals, so they infer
as Literal[4]? — an inferred literal that falls back to int — rather than a declared Literal[4].
.version therefore no longer narrows an IPv4Network | IPv6Network union, which it does on 3.13
where these are Literal-returning properties, and that narrowing is the only type-safe way to reach
subnet_of / supernet_of / address_exclude, since they take other: Self.

Annotate the literal types explicitly, with # noqa: PYI064 because the bare Final that rule
suggests is precisely the form that loses the declared literal (pre-commit.ci autofixed it away in
0752a11, reverted in 44bc4e1). Happy to drop Final and use plain Literal[4] instead if you
would rather not carry the noqas, or to take this to ruff as a rule bug first.

Agent used: Claude Code (Claude Opus 5).

---

2026-08-31 (@srittau): Deferred pending discussion in python/typing#2351.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[srittau]

: Final = X is the pattern we generally use in typeshed. The typing spec says:

The typechecker should apply its usual type inference mechanisms to determine the type of ID (here, likely, int).

This text has been copied verbatim from PEP 591 This was released around the same the PEP 586 (Literals) was released, so this might simply be a case where two typing PEPs overlapped. This is supported by the following quote from the spec, a bit later:

Type checkers should treat uses of a final name that was initialized with a literal as if it was replaced by the literal.

(Also coming from PEP 591.)

The conformance tests also ensure that.

I've opened python/typing#2351, so let's wait for the feedback there.