nx_likeness_lib.nx

nx_likeness_lib.nx -- THE IMAGE-LEVEL LIKENESS REFEREE (aesthetictwin AT45, 2026-09-18). The board had no ruler that reads a RENDER against the PHOTOGRAPHS of the person it claims to be, so a render the operator called an ugly ventriloquist dummy was published beside its reference with every gate green. This lib is that ruler, and every layer of the feedback loop above it (fit, morph, texture, hair) is judged by it. It COMPOSES the estate's incumbents and re-implements none of them: nx_photomark_lib decodes a JPEG or PNG through one door (pm_decode_file), places the four canthi and the five mouth landmarks (pm_landmarks_cascade, pm_landmarks), measures the roll-corrected canthal tilts, and parses the OUTSIDE oracle rows (pm_field, pm_span_int) the intake lane already banks per reference still nx_visdesc the grayscale edge-orientation layout descriptor (80 integers) and its L1 distance -- the STRUCTURE axis nx_cosine_similarity the signed Q10 cosine over the same two vectors -- structure read a second way nx_skinsample_lib the skin chroma rule and the linear-light mean converted to CIE L*a*b* -- the TONE axis THE SHAPE OF THE VERDICT. Two images are ALIGNED into one canvas by the similarity that maps each face's two eye centres onto the canvas's two eye centres (the same normalisation on both sides, so a roll or a scale in either still is undone before anything is compared), then SEVEN AXES are measured between the aligned pair and each is published on its own: geom_nme_permil mean landmark distance in the canvas, permil of the canvas eye span (the four canthi, plus the five mouth landmarks when both images carry a mouth); needs both images landmarked by a passing eye pair tilt_delta_deg10 the larger of the two roll-corrected canthal-tilt differences; needs both tilts measured struct_l1 nx_visdesc_l1 between the two aligned crops' descriptors struct_cos_dist NX_COSINE_Q minus the Q10 cosine of the same descriptors (0 = parallel) tone_dl_e3, tone_da_e3, tone_db_e3 the absolute L*, a*, b* differences (x1000) of the skin-masked linear mean of each crop; need skin pixels in both crops THE SAME-PERSON BAND is not a number anyone typed: `lk_band` measures every axis on EVERY PAIR of reference stills of the one person and the band per axis is the WORST (largest) distance any same-person pair reached. A subject passes an axis iff its distance to its NEAREST reference is within that band; the verdict is the count of axes inside, outside and unmeasured, and INSIDE only when no measured axis is outside (the MIN-headline doctrine: no axis may be compensated by another, so there is no weighted sum and no coefficient). The band prints the pair that set it and how many pairs derived it. EVERY NUMBER ANNOUNCES ITS RESOURCE. The canvas side is TWICE THE SMALLEST REFERENCE EYE SPAN, floored to the descriptor's grid (no reference is ever upsampled; a subject smaller than that is announced as upsampled); the eye line sits at 340 permil of the canvas height because on the anchor oracle row (ri_diora_baird_14: brow top 127.8 px, eye line 197.3 px, eye span 103.4 px) the brow top sits 680 permil of the eye span above the eye line, so with a canvas two spans wide the brow top lands on the canvas's top edge and the bottom edge falls 1.32 spans below the eye line -- past the stomion (1.15 spans on the same row) and above the chin (1.65). The smallest canvas is VD_GRID cells of three pixels, the least a Sobel neighbourhood needs. ALIGNMENT SOURCES ARE NAMED, NEVER SILENT. A still with an oracle row (knowledge/refintake/<slug>.oracle, matched by basename) aligns on the oracle's canthi at the oracle's tenth-pixel precision; a subject without one runs the photo ruler's cascade route, then its skin route, and when the skin route REFUSES its eye pair but still holds candidate canthi, those candidates align the crop while the geometry and tilt axes read UNMEASURED (the pair was refused, so its landmarks are not evidence). MEASURED the day this was written: on the rejected render diora_baird_close.png the cascade's face box was the nose tip and no eye pair followed, and the route chooser never reaches the skin route on that refusal -- a subject the ruler cannot align reads UNMEASURED, which is neither INSIDE nor OUTSIDE, so the ladder exists to turn "could not look" into a named, degraded look wherever one is possible. Sides follow the image: R is the image-left eye (the subject's right), as the photo ruler and the oracle rows agree. LIB, no main. license_tier: ORIGINAL No hw writes (Rule 26).