nx_huffcdic.nx

nx_huffcdic.nx -- SOVEREIGN HUFF/CDIC decompressor for MOBI/AZW (compression type 17480), the Nishi capability equivalent of Calibre's huffcdic reference. CLEAN-ROOM: written from the documented algorithm (HUFF/CDIC record layout + the 32-bit Huffman-window decode), NOT copied -- pure NishiLang (nx_cc -> nxasm_x86, no python/gcc/sh). Newer Kindle books use HUFF/CDIC instead of PalmDOC; nx_mobi_book currently handles 1/2 (none/PalmDOC) only. HUFF record: "HUFF\x00\x00\x00\x18" | off1(u32) | off2(u32) | dict1[256]u32 (at off1) | dict2[64]u32 (at off2) dict1 entry v: codelen = v&0x1f ; term = v&0x80 ; maxcode = ((v>>8)+1 << (32-codelen)) - 1 dict2: 32 (mincode,maxcode) pairs -> mincode[cl] = raw<<(32-cl) ; maxcode[cl] = ((raw+1)<<(32-cl))-1, cl=1..32 CDIC record: "CDIC\x00\x00\x00\x10" | phrases(u32) | bits(u32) | u16 offset table -> phrases phrase j: off = be16(16+j*2) ; blen = be16(16+off) ; bytes = [18+off .. 18+off+(blen&0x7fff)) ; flag = blen&0x8000 unpack: 32-bit window; codelen/term/maxcode = dict1[code>>24]; if !term walk mincode[] up; r=(maxcode-code)>>(32-codelen) is the dictionary index; a non-terminal phrase is itself a compressed slice -> recurse, then cache as terminal. NXASM-SAFE: be64 builds the u64 via a <<8 loop (immediate shifts >=32 are mis-encoded); every <<(32-cl) is a VARIABLE (register) shift, which x86 SHR/SHL handle correctly. license_tier: ORIGINAL