nx_img_scale.nx

nx_img_scale.nx -- sovereign image resampler, hardware-rung-up. Bilinear resize of an N-channel u8 image: ANY src WxH -> ANY dst WxH (both UP- and DOWN-scale, the general "convert to size Y" primitive). Built from the hardware rung UP: pure i64 Q16.16 fixed-point -- no float, no libc, no SIMD dependency, only integer multiply / add / divide-by-power-of-two. All intermediates are kept NON-NEGATIVE (source coord clamped into range) so the code never relies on signed-shift semantics. Convention: half-pixel centers, sx = (x+0.5)*sw/dw - 0.5, edge pixels repeated -- the SAME convention OpenCV INTER_LINEAR / Pillow / ffmpeg-swscale use, so the planned differential gate can prove byte-parity vs a 3rd-party reference (non-novel-capability proving doctrine). nx_scale.nx is Gaussian-pyramid DOWN-scale only; this is the general up/down resampler the media convert path needs. RUNG 1 (this file): bilinear, correctness via KAT (hand-computed from the standard formula = the independent oracle). NEXT: (a) differential gate vs swscale/Pillow; (b) bicubic / Lanczos as selectable vtable transforms; (c) wire into nx_uxf_negotiate's convert path; (d) MEASURED S-class exceed census (organ-graded, never self-scored). license_tier: ORIGINAL