nx_bluenoise.nx

nx_bluenoise.nx -- SOVEREIGN BLUE-NOISE MASK GENERATOR (Ulichney void-and-cluster, integer-exact). WHY THIS EXISTS: quantizing a photograph or a page of text to the ~16 grey levels an e-ink panel actually has is what produces the "ink on paper" look. Doing that quantization with a THRESHOLD MASK decides whether the result reads as paper texture or as a visible screen-door pattern. A blue-noise mask is the one whose thresholded pattern has an isotropic power spectrum with no low-frequency energy -- the eye's own low-pass then averages it to the right tone and never resolves structure. White noise clumps; ordered (Bayer) masks show a lattice. Blue noise is the state of the art and has been since Ulichney 1993. INTEGER-EXACT BY CONSTRUCTION: no float anywhere. The Gaussian is a 33-entry fixed-point table indexed by SQUARED distance, so the filter is exact and the same on every machine -- a mask that differed per host would make every downstream render non-reproducible. COST, STATED: the classical algorithm is O(N^2) in cell count. 64x64 is ~16.7M ops (fast); the published figure for 8192x8192 is about a MONTH, which is why the right shape is "generate a small mask ONCE, ship it as data, tile it" rather than generating at render time. The 128 cap below is that decision made explicit rather than discovered by a hang. IMPRECISION I CHOSE TO LIVE WITH (documented so the next reader does not trust it as exact): Ulichney's phases 3 and 4 reverse which class is the "minority" for filtering purposes. Because the filter is linear and toroidal, energy_of_zeros = constant - energy_of_ones, so "tightest cluster of 0s" and "largest void of 1s" select the SAME cell. The two phases therefore collapse into one loop here, exactly rather than approximately. nx_bluenoise gen <size> <out.raw> -- write size*size bytes, cell value = rank scaled to 0..255 nx_bluenoise stats <in.raw> <size> <fill> -- nearest-neighbour stats of the pattern thresholded at fill permil nx_bluenoise selftest -- the gate license_tier: ORIGINAL