Screen PDS endpoint hosts without Network.framework

[germ-mark]

ServiceEndpoint.swift imported Darwin and Network, so the package does not
build on Linux or Android. This replaces those with libc plus one reimplemented
parser, ahead of adding the Linux and Android CI legs.

Why the parsers couldn't just move to inet_pton

Network's IPv4Address is not a strict dotted-quad parser. Readings taken from
it on macOS:

host	IPv4Address	inet_pton	inet_aton
16843009	1.1.1.1	—	1.1.1.1
0x7f000001	127.0.0.1	—	127.0.0.1
127.1	127.0.0.1	—	127.0.0.1
0177.0.0.1	177.0.0.1	177.0.0.1	127.0.0.1
010	0.0.0.8	—	0.0.0.8

https://16843009 is an accepted endpoint in the existing tests, and it is only
accepted because IPv4Address vouches for it as a literal — inet_pton rejects
the short forms, so swapping it in would have started rejecting valid dotless
literals.

The probes pin down IPv4Address's actual shape: for a dotted quad it uses a
permissive grammar of its own (leading zero decimal, 0x hex); for everything
else it defers to inet_aton, octal included — 010 is 0.0.0.8, and
018015111 is no address at all (8 is not an octal digit). IPLiteral.v4
mirrors exactly that: a reimplementation of the quad grammar, pinned by tests,
and legacyV4/inet_aton for every other shape. v6 moves to
inet_pton(AF_INET6) — byte-identical to IPv6Address on every case probed,
including embedded v4 and %zone suffixes — and IPLiteral.embeddedV4
reproduces asIPv4, since the ::1 distinction is what the loopback exemption
turns on.

Getting the non-quad deferral right matters for the screening, not just
compatibility: a decimal-everywhere reading would have accepted dotless
spellings like 018015111 as public literals, exempting them from the
single-label rule that exists because such names resolve through local search
domains. https://018015111 is now pinned as rejected in PDSEndpointTests.

Behavior

Unchanged on Apple platforms — all 33 existing tests pass untouched. The one
probed divergence from IPv4Address is a quad part of 0x with no digits
(1.0x.2.3), which read as zero there and has no reading here; it loses the
literal exemption and is screened as a name, and where the platform's
inet_aton reads it, that reading is still screened.

IPLiteralTests pins the accepted spellings (quad-decimal and octal-deferral
forms), the rejected ones, the leading-zero disagreement between the two v4
readings, and the v4-in-v6 table against readings probed from the real Network
parsers, so drift in the reimplementation fails a test rather than opening a
hole.

Part of GER-2175.

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: 17738ad

The changes in this PR will be included in the next version bump.

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR