nx_vcodec_compound_ceil.nx

nx_vcodec_compound_ceil.nx -- GEN-2 RUNG-0 PROBE P0a: compound-from-past AVERAGING ceiling (SAD domain). Question: on P-frames, does predicting a coded MB as the AVERAGE of two motion-compensated PAST references (latency-safe: no future frame) beat the single-ref prediction we ship? Compound averaging cancels independent quant noise between refs (up to ~sqrt(2) on static) and covers occlusion flips -- multi-ref SELECTION measured only 1.5%, but AVERAGING is unmeasured and is gen-2's biggest open lever. Method: build the recon chain with the REAL shipped encoder (vv_enc_rct8, emode 681); per coded (non-skip-threshold) MB, compare full-pel ME SAD: single = vm_search vs prev1; compound = avg(MC(prev1,mv1), MC(prev2,mv2)) with mv2 = vm_search vs prev2. Both arms full-pel (equal granularity = fair relative). Reports total SADs + the per-mille cut + the share of MBs where compound wins. ⚠SAD->bits is SUBLINEAR (measured: an 18% SAD cut -> 3.7% bits at qp8) -- read the result as an UPPER-bound direction gate for building compound syntax, not as BD-rate. license: ORIGINAL