The storage primitive: bit-packed unsigned buffers
Every integer dataset is the same primitive: numEntries values of width bits, packed MSB-first into big-endian bytes
width is the minimum multiple of 8 that fits the value range (MinBits of the max id), so a store with 40M objects spends 4 bytes per object id, not 8
1-bit width = the B bitmaps; 64-bit width carries full two’s-complement longs
Byte-aligned widths are what make the parallel writers possible: entry i occupies exactly bytes [i*w/8, (i+1)*w/8) — concurrent positional writes never share a byte
Readers address entries by long offsets (no 2 GiB ceiling) over jHDF-mapped, FFM-mapped, or HTTP-ranged bytes
binarySearch / lowerBound / upperBound operate directly on the packed form with unsigned comparison