nx_vcodec_stage_bench_recovery_20260912.nx

nx_vcodec_stage_bench.nx -- WHERE DOES THE ENCODE TIME GO. Stage attribution of the LIVE rct8 RTC encoder, per macroblock, on the REAL head-to-head test card (the akiyo index-strip card nx_video_h2h judged 2026-09-02). WHY IT EXISTS: the same-instrument head-to-head measured ours at 9.9-18.2 of 30 fps against Jitsi's 29, and the probes' own arming dump showed the cause is encode COST (93 ms single-thread, 48 ms banded at 640x480), not the ladder (30/45/60 rungs exist) and not wasm (native 74 us per MB == browser 77 us per MB). Nothing in the estate attributed that cost to a stage (nx_capsearch 2026-09-02: no encoder stage profiler), so every fix would have been a guess. This organ MEASURES: (A) the whole production frame encode, (B) the integer motion search kernel called once per MB exactly as the block coder calls it, (C) the quarter-pel refine kernel on the CODED MBs (and on all MBs, the ceiling), and derives REST = A - B - C (transform, quant, RDOQ, entropy, recon, deblock). Kernel costs are the cost of ONE call per MB, which is an UPPER BOUND on their share if the encoder early-outs before them. Historical baseline: emode 2593 = range coder + sig-map + gentle deblock + RD-skip. Build 865 additionally enables quiet-quadrant skip (emode 18977); both arms remain in this benchmark. The live skip threshold is qp*30; qp*188 is the separate RD-benchmark control. Two qp points bracket the ladder. Prints one line per stage per qp; last line is the verdict for gv_last_line readers. Card: 352x288, first 120 frames of the h2h card (banked in _ops/video_h2h/testcard120.y4m). license_tier: ORIGINAL