nx_zstd_huf_gate.nx

nx_zstd_huf_gate.nx -- proves the zstd Huffman decode table. T2 pins the EXACT table layout for a known weight set, index by index. Weights {2,1,4,3,1} at tableLog 4 must tile 16 entries as: 0..7 symbol 2, 1 bit (weight 4, the shortest code, LOWEST indices) 8..11 symbol 3, 2 bits 12..13 symbol 0, 3 bits 14 symbol 1, 4 bits 15 symbol 4, 4 bits Laying ranks out in SYMBOL order instead of DESCENDING WEIGHT produces a table that decodes self-consistently and disagrees with every other zstd implementation -- a file only this decoder can read. Asserting the indices directly is the only way to catch that. T3 requires the slots to TILE EXACTLY. A weight set that sums to 15 of 16 leaves an entry no symbol ever claims, and a peek landing there would decode as symbol 0 forever. The build refuses rather than leaving the hole. T5 decodes a hand-assembled backward stream. Bits were placed so the reader emits 0 | 10 | 110 | 1111, which must come out as symbols 2, 3, 0, 4 -- four codes of four different lengths, so a decoder that consumes tableLog bits per symbol instead of the code's true length fails on the second one. license_tier: ORIGINAL