nx_attention_class.nx

nx_attention_class.nx -- sealed enum: cell attention class. Per cardinal: cooperative-resource-arbitration. The attention class is the cell's declared priority intent. Foreground game holds concurrent VRAM and CPU against background inference; idle work runs only when nothing else is asking. Distinct from nx_attention.nx (ML scaled-dot-product attention). The semantic collision is at the name prefix only; this primitive concerns OS-level resource arbitration, not transformer kernels. Composes: nx_budget -- budget grant decisions honor class priority nx_yield -- yield request from lower class is auto-granted when higher class is contending nx_evict_journal -- evictions log displacing class for forensics nx_homeostasis -- migration tropism preferences vary by class (foreground stays local; idle is offload-OK) V1 ships seven classes per cardinal. Priority order is FIXED here so substrate-wide behavior remains reproducible across hardware tiers. Gap list (V1 honest perf verdict): - no nested classes (cell has exactly one class at a time) - no temporary boost (priority does not change on key-press; callers must re-instantiate cell with new class) - no fairness within same class (FIFO on contention, not weighted) genealogy_id: nishi_cardinal_2026-05-17_cooperative_resource_arbitration lineage_id: substrate_attention_class_v1 nx_safety_envelope: intended_use: "Declared cell priority class for resource arbitration; honored by budget + yield layers" sil_target: SIL1 evidence: [enum_sealed, priority_total_order] verdict: NOT_YET_EVALUATED