nx_extllm_lib.nx

nx_extllm_lib.nx -- THE EGRESS DECISION LIBRARY: may THESE BYTES leave, to THIS PROVIDER, for THIS USE? The CODE half of two conf files authored 2026-08-24 that have had no reader since (measured 2026-08-25: xl_egress matches=0 over 61,475 files, coverage_complete=1 corpus_complete=1). xl_class_of -- CLASSIFY AT THE DOOR over the operator's release manifest. C by default; U only on an explicit row whose sha256 still matches the file. A STALE row admits NOTHING. xl_use_verdict -- the provider terms ledger as data. UNKNOWN is a VALUE and REFUSES every gated use. xl_budget_* -- CUMULATIVE disclosure per provider. Splitting work into parts defeats a SINGLE call and NOT aggregation across calls: a provider holding a perpetual licence retains every fragment and reassembles at leisure. The bound is on the SUM, never on the call. WHY A NEW LIB RATHER THAN EXTENDING AN INCUMBENT (checked before building, 2026-08-25): the estate's three shipping classifiers cannot answer an OBJECT-side release question. ag_uid_to_level grades a READER (a 0..3 RBAC ladder, never a byte's sensitivity); nx_site_lock_lib slk_resolve returns 0 for an unmatched path and ag_allow turns that into ALLOW for everyone ("no match => PUBLIC"); the docportal consent bit is OPT-OUT (keep initialises to 1, so an undeclared document is SERVED). Two of the three default to release, so this decision must fail closed INDEPENDENTLY of all of them rather than compose one. Verdicts are exit-code shaped so a caller can branch without parsing. license_tier: ORIGINAL. No hw writes (Rule 26).