nx_vcodec_hpel6_probe.nx

nx_vcodec_hpel6_probe.nx -- 6-TAP HALF-PEL SIZING (derived from entcost probe) (seq1105, the 2.9x direction question): is the residual coder spending materially more bits than its own coefficients' entropy floor? For each P-frame MB of a real 2593-chain encode (same refs the encoder used), take the encoder's deterministic (mv,qpel) choice, form the residual, 4x4-transform+quant (the shipped tf=1 path), then measure: bits_rc = rc_sig_cost_q8 (the shipping coder's own cost) vs H0 = the static per-zigzag-position level-entropy floor of the WHOLE collected coefficient field. bits_rc / H0 <= ~1.15 => the coder is fine, the gap is UPSTREAM (prediction/transform quality); bits_rc / H0 >= ~1.3 => context/entropy depth is the rung. No float: entropy accumulated in millibits via a log2 lookup on p in permille. license: ORIGINAL