nx_source_health.nx

nx_source_health.nx -- SOURCE HEALTH, THE DECISION + THE PERSISTED VERDICT (/compare/mediaingest R1; contract symbol sh_preflight). LIB: pure classification + a tiny per-host verdict table. The heavy I/O probe (fetch + bot-wall classify) lives in the thin program nx_source_health_probe so THIS lib's import closure stays small (nx_syscalls only) and the gate can prove every branch with planted numbers -- the same lib/program split that keeps a gate honest and dodges the nx_cc module-limit that a fetch+antibot closure would hit here. WHY: Harvestr fast-fails a null-routed CDN (its April-2026 Coomer route-around) before queueing work against it; ours had nx_site_liveness for OUR sites only. This is the general per-host pre-flight: probe once, PERSIST the verdict, and let every capture read it so a dead host is skipped WITHOUT re-probing on every asset. FIVE STATES (a host that cannot be classified must not read as UP -- abstain, never flatter): SH_UP the host answered (any 2xx/3xx, or a 4xx/5xx where the HOST is alive but the path is bad) SH_SLOW it answered but slower than the slow bar -- usable, flagged SH_DEAD connect failed / timed out -- null-routed or down; skip and say so SH_WALLED it answered with a bot-wall (nx_antibot verdict) -- classified, never solved SH_UNKNOWN never probed / stale -- the third state, so "no evidence" is not "healthy" license_tier: ORIGINAL