nx_view_transform_lib.nx

nx_view_transform_lib.nx -- THE ONE OWNER OF THE ESTATE VIEW TRANSFORM (GR23, 2026-08-28). license_tier: ORIGINAL No hw writes (Rule 26). WHY THIS EXISTS. The shipped beach page carries the value-structure grade INLINED FOUR TIMES (world FS, figures, hair, WGSL) as a hand-typed .94*(x/.94)^1.45, derived offline in Python and reproducible by nothing. Operator directive 2026-08-28: "we dont want webgl2 and other 3rd party unless we are recreating ourselves the capabilities". So the curve is DERIVED HERE, in NishiLang, from recorded measurements, and BOTH execution surfaces are emitted from ONE generated LUT so they cannot diverge. AgX and ACES are ORACLES we check ourselves against, never a dependency -- borrow the rulers, not the runtime. Neither appears in this file as code. NOTHING BELOW IS TYPED FROM A CALCULATOR. The only numbers are (a) the measurements, declared as data with their provenance, and (b) values DERIVED at runtime. The exponent is computed, not typed; the LUT size is searched, not chosen. THE MEASUREMENT (PIL over the operator own 20:19 frame vs the reference photographs, 2026-08-28): saturation is NOT the gap (S mean .29 against a reference band .27-.32, lift x1.01); the gaps are HUE MONOCULTURE and HIGH-KEY MIDS (V p50 .79 against a reference band .59-.66). THE CURVE. A p95-anchored power pivot: out = k*(x/k)^g, k the frame own p95 so the white point is an exact fixed point. g is fixed by requiring the median to move a declared fraction of the way from where it is to the reference band near edge: m_tgt = m_src - (m_src - band_hi) * num/den g = ln(m_tgt/k) / ln(m_src/k) HALF-way (num/den = 1/2) is the declared choice, because sand content legitimately differs from the reference scenes. MEASURED RESIDUAL: this yields g = 1.4939, while the shipped page carries a hand-typed 1.45, which moves the median 45.7% -- not the "half" its own comment claims. A constant that has drifted from the rule written beside it is exactly what a generator removes. THE APPLICATION IS LUMINANCE-SCALED, AND THAT IS THE FIX, NOT A REFINEMENT. Applying a power per-channel raises saturation BY CONSTRUCTION: with r = min/max, HSV S = 1 - r becomes 1 - r^g, and g > 1 forces r^g < r, so S rises for every non-grey pixel, monotonically in g. Measured on the live frame the shipped per-channel form drove S mean .313 -> .411, i.e. OUT of the .27-.32 band it was told saturation already sat inside. Scaling the triple by the ratio the curve applies to its LUMINANCE leaves min/max untouched, so S is preserved EXACTLY, for every pixel, by construction. Measured: V p50 .729 -> .616 (inside .59-.66) with S mean .313 -> .313 (unchanged). Both bands held. DECLARED EXPOSURE, NOT HIDDEN. The curve crushes deep shadow: over inputs 0..0.10, 26 distinct 8-bit levels collapse to 10 (38.5% survive), and 5 of 256 input levels map above 1.0 and clip. That is invisible on a noon beach and is where a night or interior frame will show banding. It is