Constant MODEL_HOSTS 

pub const MODEL_HOSTS: &[&str];

Known machine-learning MODEL-provider hosts — the SPEC §1 ⟨0.13⟩ Llm host-literal refinement: a statically-known Net request to one of these classifies Llm IN ADDITION to Net (Net is never dropped — a model call IS network I/O, exactly as an Exec-refined subprocess keeps Exec), just as a jdbc URL classifies Db. Matched by host, case-insensitive; a SUBDOMAIN of a listed host counts. The reference engines share this table VERBATIM with candor-java’s Literals.MODEL_HOSTS (the analog of classify_command_head) so the Net boundary refines to Llm identically. An UNKNOWN host stays bare Net — never guessed. Curated STARTER set; the §7 coverage ledger discloses an uncovered provider like any other.