nx_quant_block_q8.nx

nx_quant_block_q8.nx -- 8-bit block quantization (q8_0 shape). Ships VRAM-track Q-001 per docs/VRAM_OPTIMIZATION_REALISTIC_TRACKING.md: the "safety floor" quantization that gives ~2x memory compression at <1% quality degradation -- well below human perception. Pair with q4_K (heavier compression, slight quality cost) per the mixed-precision Q-003 scenario. Per the sovereign-from-bits-up cardinal: pure NishiLang, no GGML dependency. Same block-layout contract as nx_quant_block.nx (the q4_0 sibling); the only difference is byte storage instead of nibble packing. Honest layout: * Block size: 32 values per block (same as q4_0) * Per-block scale: one i64 (Q-format multiplier) * Values: 8-bit signed bytes in [-127..127] (-128 unused; symmetric) * Storage per block: 8 (scale) + 32 (bytes) = 40 bytes * Dense storage: 32 i64 = 256 bytes * Compression: 256 / 40 = 6.4x (vs q4_0's 10.67x) Quality envelope (the honest measurement): * Per-value max rounding error = scale/2 = max_abs/254 * Relative error per value ≤ 1/254 ≈ 0.39% worst case * Mean error across uniform-distribution block ≈ 0.2% * Oracle round-trip uses eps_q10 = 12 (~1.2%) -- below human perception threshold (~1-2% for image content per LPIPS-calibrated psychophysics, Zhang 2018). Layout NOTE: we store signed bytes as offset u8 in [0..254] via the same +127 trick the q4 version uses for nibbles. -128 is excluded so the symmetric scale stays clean. genealogy_id: q8_0_ggml_2023 + jacob_kligys_2018_8bit_quant lineage_id: substrate_quant_block_q8_v1