nx_mrshuffle.nx

nx_mrshuffle.nx -- the real MapReduce SHUFFLE (closes distributed-compute's shuffle-exchange gap). The hard, distinctive part of distributed group-by/join is the SHUFFLE: repartition rows by a hash of the key so that every row for a given key lands at exactly ONE reducer, and each reducer sees ALL rows for its keys -- then reducers aggregate independently and their union is the answer. This is the canonical COUNTING shuffle (count per reducer -> prefix-sum to contiguous offsets -> place), so it is O(n) not the O(n*R) broadcast-filter shortcut, and the reduce phase runs in REAL parallel (one forked process per reducer, over a shared exchange). From the first byte up: fork + shared mmap + wait4, no frameworks. Composes the zero-copy nx_colframe. license_tier: ORIGINAL No hardware writes (Rule 26).