nx_quality_raci_test.nx

nx_quality_raci_test.nx -- proves the feature is ROBUST and correctly DISTRIBUTED across organs: ENGINEER diagnoses the bound, BUILDER decides the operating point under the COUNCIL's quality policy. Exit 0 only if every check holds. (1) decode matvec (Q4) -> Engineer: MEMORY-bound, lever=DATA. (2) Builder, balancing quality(slightly>speed) under the Council floor, picks Q4_K_M -- which the published data independently calls the best quality/size balance. NOT the fastest. (3) contrast: a pure-SPEED choice would pick Q2_K (smallest, +8.82% perplexity) -> proves the balance is doing real work, and quality is being protected. (4) compute-bound kernel -> Builder keeps HIGHEST quality (bytes don't bind; no quality sacrifice). (5) COUNCIL floor rejects Q2_K/Q3_K_M, admits Q4_K_M. (6) ROBUSTNESS: invalid input (0 bytes) -> RF_UNKNOWN (no silent verdict); a 4-exabyte kernel (overflow range) still returns a valid bound; all-below-floor -> graceful highest-quality. license_tier: ORIGINAL