nx_ribosome.nx

nx_ribosome.nx -- target-ISA decision metadata. Biological analogue: a ribosome translates messenger RNA into the specific protein its codons encode. Nishi ribosome is the NishiLang-side decision primitive that names WHICH of the existing nxc2 codegen backends should emit a call site's binary. The actual codegen lives in C (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). This primitive's job is to say "this site -> emit AARCH64 NEON" or "this site -> ESP32 Xtensa" based on observed heat from nx_metabolism + the cell's tropism. THIS IS WHAT MAKES DOOM-ON-PREGNANCY-TEST PHYSICALLY POSSIBLE: the same .nx source compiles down to CORTEX_M Thumb on $5 silicon (T1) and to PTX on a 5090 (T5). One IR, fourteen targets, decided per call site by ribosome. Composes: nx_tier -- NX_TIER_MCU..NX_TIER_HPC informs target choice nx_metabolism -- hot sites get SIMD/GPU/SILICON targets nx_pathway -- cell-spec declares preferred ribosome target nx_tropism -- LOCAL_GPU tropism selects GPU-family targets V1 ships a sealed enum + a picker. The handoff to nxc2 is via the target tag on each call site; the nxc2 main loop already reads per-site target hints (see nxc2/main.c). What was missing in NishiLang was the decision primitive; here it is. Gap list (V1 honest perf verdict): - SILICON_HDL emit target is a stub (HDL backend queued) - no online learning -- picker is fully deterministic - no per-cell override (uses pathway-level declaration) - no cross-target ABI shim (caller must keep i64/i32 alignment constant; nx_tier's nx_int alias is the single tuning knob) genealogy_id: nishi_cardinal_2026-05-17_pathway_tropism + biology_protein_synthesis lineage_id: substrate_ribosome_v1 nx_safety_envelope: intended_use: "Per-call-site target ISA decision metadata