nx_eink_tone.nx

nx_eink_tone.nx -- PERCEPTUAL TONE QUANTIZER for e-ink-style rendering. Integer-exact, no float. WHY: an e-ink panel has ~16 grey levels, not 256. Getting from a photograph or a rendered page down to 16 levels is where the "ink on paper" look is won or lost, and it turns on TWO decisions that are usually got wrong: 1. QUANTIZE IN PERCEPTUAL LIGHTNESS, NOT IN sRGB OR LINEAR. sRGB code values are not perceptually spaced, so 16 evenly-spaced sRGB steps put most of their resolution where the eye cannot use it and band visibly in the shadows. CIE L* is defined to be perceptually uniform, so 16 evenly-spaced L* steps look evenly spaced. This is the difference between "16 greys" and "16 USEFUL greys". 2. DITHER WITH A BLUE-NOISE MASK, NOT BY ROUNDING. Rounding to the nearest level produces flat bands. A blue-noise threshold (nx_bluenoise) breaks the band into a texture the eye integrates back to the true tone without ever resolving structure. The sRGB->linear table is BAKED DATA, computed from the exact piecewise sRGB transfer function and scaled to 1e5. It is a table rather than a runtime power because x^2.4 in integer arithmetic is either wrong or expensive, and a transfer curve that differs per host makes every render irreproducible. KNOWN-ANSWER ANCHOR: sRGB 128 -> linear 21586 -> L* 53.59, the textbook value for mid-grey. The gate asserts it, so a silent regression in the table or the cube root cannot pass. nx_eink_tone lstar <r> <g> <b> -- CIE L* x1000 for one sRGB colour nx_eink_tone ramp <levels> -- the quantization ladder, in L* and in sRGB nx_eink_tone dither <w> <h> <gray> <levels> <mask> <masksize> -- dither a FLAT tone, report the mix nx_eink_tone selftest -- the gate license_tier: ORIGINAL