nx_quality_balance.nx

nx_quality_balance.nx -- BALANCE quality against speed (the operator's law: always balance quality and quantity, with QUALITY SLIGHTLY MORE VALUABLE than raw speed). Cutting bytes (lower quantization) buys speed in the memory-bound regime but costs QUALITY (perplexity rises). "Just go Q2 for 1.8x" is wrong -- Q2_K is +8.82% perplexity. This module picks the operating point that maximizes a quality-weighted score under a hard QUALITY FLOOR (rule: never strip quality for speed). Data-driven: the quality + size numbers are measured config, not magic. RACI: the COUNCIL governs the POLICY here (the quality floor + the quality:speed weighting -- the value judgment). The ENGINEER supplies the measured numbers (nx_engineer_profile). The BUILDER consumes qb_select to choose the operating point (nx_builder_quant). One organ per job. Real llama.cpp data (Llama-3.1-8B class), perplexity delta to fp16 as PERMIL (pct*10), and bits per weight *10: fp16(0,160) Q8_0(1,85) Q6_K(2,66) Q5_K_M(4,55) Q4_K_M(12,45) Q3_K_M(47,34) Q2_K(88,26). Refs: arxiv 2601.14277 unified quant eval; llama.cpp perplexity discussion #406. license_tier: ORIGINAL