nx_vcodec_entropy.nx

nx_vcodec_entropy.nx -- R2 (build the codec UP, never settle): the flagged "rANS replaces RLE" entropy upgrade, MEASURED on REAL transform coefficients. The codec's lossy gap to x264 (~2.05x) is partly the entropy stage: fixed run/size RLE codes leave bits on the table vs an adaptive arithmetic coder that spends -log2(p) bits per symbol. We REUSE the proven sovereign range coder (nx_range_coder, patent-free arithmetic coding -- NOT CABAC) with an adaptive frequency model over the (run,size) tokens of real 4x4-WHT-quantized coefficients, mantissa carried as an explicit field. MEASURED: adaptive-arithmetic bytes vs fixed-RLE bytes on a real-ish frame; the entropy stage is LOSSLESS (decoded coefficients == encoded, bit-exact). Neg-control (TAMPER): corrupt a stream byte -> the decoded coefficients MUST change (the bytes are load-bearing -- the win is not fabricated). Positioned vs the x264 yardstick: prior gap 2.05x -> 2.05 * (rc/base) after this stage. HONEST SCOPE: improves ONLY the entropy stage; RDO + trellis-quant + in-loop deblock + the neural path are the bigger remaining levers to actually reach/exceed x264. main() is the SELF-VALIDATING GATE. Evidence -> knowledge/status/vcodec_entropy.log. license_tier: ORIGINAL