nx_metabolism.nx

nx_metabolism.nx -- per-call-site profiler + tier promotion. Biological analogue: cells consume more glucose at hot signaling pathways and grow mitochondria to compensate. Nishi metabolism tracks (hit_count, total_cycles, ram_peak) per call site and suggests a target tier for the existing nxc2 codegen layer to emit: interpreted bytecode -> AOT native -> SIMD-vectorized native -> GPU offload -> custom-silicon block. THIS IS THE OPTIMIZATION DECISION LAYER. nxc2 already has thirteen architecture backends compiled in (aarch64.c, armv7a.c, cortex_m.c, gcn.c, loongarch64.c, metal.c, mips64.c, ppc64le.c, ptx.c, riscv.c, riscv32.c, s390x.c, wasm.c, x86_64.c). What was missing is the NishiLang-side decision primitive that says WHICH backend to pick per call site. Metabolism is that primitive. Composes: nx_tier -- NX_TIER_MCU..NX_TIER_HPC target enum nx_budget -- ram_peak per site informs ceiling tuning nx_homeostasis -- repeated hot-site overshoot triggers tropism migration to higher tier nx_attention_class -- foreground sites get higher promotion bias (frame-budget pressure outranks build pressure) V1 ships a flat hash table keyed on site_id. Adaptive thresholds per tier; defaults derived from heuristics rather than learned. Gap list (V1 honest perf verdict): - no online learning (thresholds are static constants) - no decay (a one-time spike at startup never ages out) - no cross-site cost-amortization (each site decides alone) - no silicon target yet (NX_TIER_SOVEREIGN_CHIP returns same as HPC) genealogy_id: nishi_cardinal_2026-05-17_pathway_tropism + cardinal-21_resource_awareness lineage_id: substrate_metabolism_v1 nx_safety_envelope: intended_use: "Per-call-site profiling and target-tier promotion suggestions for the codegen layer" sil_target: SIL1