_codec_real_med.nx

_codec_real_med.nx -- X-CDC-FRAME-LZ-001 representative-corpus rung: the SAME predictor-isolation head-to-head as _codec_bench_med, but on REAL PHOTO pixels (a decoded gallery PNG) instead of synthetic frames -- closing the "synthetic is unrepresentative" caveat with a no-wave verdict. Pipeline: read a real 1024x768 8-bit RGB PNG -> nx_png_decode -> raw pixels -> sample 3 deterministic 64x64 patches (1/4, center, 3/4 positions; no cherry-pick) per RGB channel -> for each plane compare MED(JPEG-LS)-prefilter+deflate vs PNG-Sub-prefilter+deflate (SAME sovereign deflate, SAME pixels). No-false-green: MED+deflate round-trips end-to-end on each plane + a tamper check that MUST fail. GREEN iff round-trip exact on all planes AND tamper detected; the win COUNT is reported, never forced. Dimensions/color are read from raw IHDR bytes (avoids the nx_int struct-field-alias hazard); only res.pixels (a pointer field at offset 8, before any nx_int field) is read from the result struct.