nx_vcodec_subpel_ceiling.nx

nx_vcodec_subpel_ceiling.nx -- RE-AUDIT the self-judged "6-tap sub-pel = 0" verdict (task #17) against a MEASURED residual ceiling, decoder-transparent. The codec's sub-pel is BILINEAR (qp_pixel: a 4-neighbour weighted blend that BLURS the reference). H.264 uses a SHARP 6-tap FIR (1,-5,20,20,-5,1)/32 that preserves high-frequency detail -> a tighter prediction -> smaller residual -> fewer bits at identical quality. This probe isolates FILTER SHARPNESS: for each interior 16x16 P-block, integer-search the clean prev (+/-16), then refine to HALF-pel with (a) the existing bilinear filter and (b) a proper 6-tap filter, and sum the best SAD each way over several P-frames. If 6-tap's total residual is materially lower, the "0" verdict was a short-bench lie (like RD-skip and wide-ME were) and the real enc+dec+wasm build is justified. Pure measurement on CLEAN originals => no bitstream, no parity risk. license_tier: ORIGINAL