Add isGlobal() and pin the classifiers to the IANA special-purpose registries

[beaugunderson]

A guard written as an OR of named classifiers (isPrivate() || isLoopback() || isLinkLocal() || …) covers only the ranges it names, and the IANA special-purpose registries hold more ranges than there are names. A differential sweep of every registry block boundary (plus the IPv4-mapped and NAT64 forms of each) against Python's ipaddress found eight ranges no classifier caught: 0.0.0.0/8, the IETF protocol assignments in 192.0.0.0/24 and 2001::/23, the three IPv4 documentation blocks, 198.18.0.0/15, 240.0.0.0/4, 100::/64 and 100:0:0:1::/64, and 3fff::/20. Each is the shape of the last four advisories: a well-formed literal that parses and classifies as unremarkable.

isGlobal() on both classes answers from a table in src/v4/constants.ts / src/v6/constants.ts that mirrors the registry row for row (block, name, Globally Reachable). The most specific containing entry with an answer wins, so 192.0.0.9 (PCP anycast, reachable) is global inside the non-reachable 192.0.0.0/24. Multicast is not global; Teredo and 6to4 (registry N/A) are not global; IPv4-mapped and NAT64 well-known forms answer for the embedded IPv4 address. SECURITY.md and the README now point guards at if (!address.isGlobal()) reject().

Parity methods: Address4.isDocumentation(), isBenchmarking(), isReserved(); Address6.isDocumentation() covers 3fff::/20 (RFC 9637) and Address6.isBenchmarking() is new; getType() learns Benchmarking, Discard-only and the second Documentation block.

Registry-pinned test: test/data/iana-corpus.json (592 probes) is generated by scripts/gen-iana-corpus.py (stdlib only; --fetch refreshes the CSVs from iana.org) from the registry CSVs checked in beside it. test/iana-corpus-test.ts asserts isGlobal() matches the registry on every probe and that isLoopback/isLinkLocal/isMulticast/isUnspecified match Python. Python 3.14 and the registry disagree on four rows (192.88.99.2, 100:0:0:1::/64, 2001:1::3, 5f00::/16); those are listed with a tripwire that fails once Python catches up. A registry change becomes a failing test after a regen instead of an inbound advisory.

Second pass, from a survey of recent advisories. Pulled every GitHub Advisory Database entry for the IP-parsing libraries across ecosystems plus every CWE-918 advisory since 2022 that mentions an address form (986 total), harvested the 525 distinct address literals they cite, and ran them through the library against Python. Every ip-CVE form (0x7F.1, 127.1, 01200034567, 012.1.2.3, 0177.0.0.1, decimal, trailing dot, whitespace, fullwidth digits) is rejected; ::fFFf:127.0.0.1, 000:0:0000::01, ::, 6to4/Teredo-embedded internals, and NAT64 local-use forms all classify as not global. The one gap was the deprecated site-local fec0::/10, which GHSA-w98g-5w9p-p3rc (2026-08-28) names as a deny-list miss: the special-purpose registry lists exceptions, but the IANA IPv6 Address Space Registry allocates only 2000::/3 for global unicast. Address6.isGlobal() is now gated on 2000::/3, which also covers the deprecated IPv4-compatible ::/96 and unallocated space like 4000::/3; the address-space CSV joins the fixtures and its boundaries join the corpus (730 probes). Python reports those reserved ranges as global, so the Python cross-check skips them with a comment.