nx_term_extract.nx

nx_term_extract.nx -- end-to-end learning orchestrator. For each beat in a StoryDb: 1. Look up the beat's source bytes from the MediaPool. 2. Run nx_noun_phrase_scan against the beat's span. 3. For each candidate, observe it in nx_term_registry (registry grows itself; no pre-seeded dictionary). 4. Back-fill beat.outfit_id / location_id / character_id with the registry slot indices for the strongest matches. Closes the loop between the ingestion (Tier 0+1) and the director pipeline: downstream nx_arc / nx_moment_resolve consume beats with populated element ids, with the actual vocabulary surfaced from the ingested corpus itself. Per cardinals: feedback-self-surfacing-intelligence: Registry stays in STAGE1 SUPERVISED today. Promotion to STAGE2 happens automatically after N sightings + 0 contradictions (per nx_term_registry policy). STAGE3 requires user pin or arc-anchor confirmation. feedback-loras-and-negatives-are-patches: Confidence is per-term auditable basis points, not an opaque embedding. Every beat-field assignment is provenance-traceable back to (beat_id, span_start, span_end, cue_index). nx_safety_envelope: intended_use: "Walk a StoryDb's beats, surface noun-phrase candidates, register them, and back-fill beat.outfit_id / location_id / character_id." sil_target: SIL2 asil_target: QM dal_target: DAL C iec_62304_class: NONE evidence: [no_floating_point, bounded_outer_beat_loop, bounded_inner_candidate_loop, registry_overflow_returns_error, no_topical_dictionary_dependency] hazard_register: [bug-tape-candidate-spans-cross-paragraph-boundary,