feat(mutators): add shuffle/unshuffle byte-transposition filter

[batmac]

Follow-up to closing #850: rather than adding a dependency on a superseded 2007 float codec, this implements the classic byte-shuffle filter that Blosc, HDF5, Parquet and Zarr all ship — in ~90 lines of pure Go, no dependency.

Why it works

An IEEE-754 double keeps its sign and exponent in the high bytes and its noisy mantissa in the low ones. For a series of similar values the high bytes barely change:

100.1 -> 66 66 66 66 66 06 59 40
100.2 -> cd cc cc cc cc 0c 59 40
100.3 -> 33 33 33 33 33 13 59 40      bytes 6-7 identical, byte 5 gradual,
100.4 -> 9a 99 99 99 99 19 59 40      bytes 0-4 noise

The redundancy is real but it lives in columns, while the file is laid out in rows. LZ77-family compressors (gzip, zstd, lz4) only find repeated sequences, and the stable bytes are stranded one per eight — never adjacent. So they find nothing and add framing overhead, which is why zstd grows float64 data.

Shuffling transposes the byte planes, turning those columns into contiguous runs:

66 cd 33 9a | 66 cc 33 99 | ... | 06 0c 13 19 | 59 59 59 59 | 40 40 40 40

Results

160000-byte inputs:

	zstd	xz	shuffle,zstd	shuffle,xz
random doubles	160019	143208	141909	129828
smooth series	160019	124732	126483	107812

Both beat what fpc achieved in #850 (143137 / 111527), with no new dependency.

Design

• Chained, not standalone: it compresses nothing by itself. -m shuffle,zstd to pack, -m unzstd,unshuffle to unpack.
• Element size is the usual X:8 argument — so shuffle:4 for float32/int32, shuffle:3 for RGB pixels, any fixed-width record with correlated fields.
• No header, still streaming: 64KiB blocks, and both directions derive the same block size from the element size, so they stay in step without metadata in the stream. Bytes that do not fill an element pass through untouched.
• Registered under the filter category, next to the other non-compressing transforms.

Verified

Round-trips byte-exactly through files and through a 4-stage stdin pipeline (-m shuffle,zstd,unzstd,unshuffle). Tests cover empty input, sub-element input, non-multiples of the element size, sizes 1/3/4/8, multi-block data and a partial final block, plus an assertion that shuffle+zstd genuinely beats zstd alone. golangci-lint v2: 0 issues; full suite green; builds on the nohl,fileonly tag set too.

🤖 Generated with Claude Code