nx_matmul_traced.nx

nx_matmul_traced.nx -- BP-3: real ML primitive (matmul) with full bit-provenance per output cell. Cardinal: feedback-no-skips-bit-provenance-silicon-to-death. Proves the user-asked pattern: every output bit of a real ML primitive is traceable to the input bits that contributed to it. A * B = C (2x2 i64 matrices) For each output cell C[i,j]: - allocate a new bit ID - emit3 the bit with up to 3 source contributions (since 2x2 needs sum of 2 products = up to 2 source layers; emit3 leaves slot for the layer-tag and primitive-tag) - the eventual consumer (display / dot / next layer) calls nx_bprov_join to close the chain expect_exit: 0