nx_vcodec_multiref_gate.nx

nx_vcodec_multiref_gate.nx -- MEASURES the latency-safe B-frame lever: does predicting from TWO PAST references (multi-ref best-of + COMPOUND averaging) shrink the residual vs single-ref (previous frame only)? This is how low-latency RTC codecs get B-frame-like gains WITHOUT the reorder/latency of classic bidirectional B (which references a FUTURE frame). Per 16x16 block, on real frames, measures total residual SAD for: (1) single-ref = prev frame N-1; (2) multi-ref = best-of {N-1, N-2, N-4 long-term}; (3) compound = average of the two best past predictions. The SAD reduction = the upper-bound prediction gain a multi-ref/ compound generation would win. GREEN reports the measured reduction (build-decision, not pass/fail-of-idea). license_tier: ORIGINAL