Targets the residual streams produced by predictor-based transforms (notably [`crate::transforms::PredictorXor`]) — those streams have a heavy distribution of small-magnitude integers with many leading zero bytes after the high-order byte. The codec encodes each value with a 4-bit leading-zero-byte count (enough to count 0..=8 zero bytes for the widest supported element) followed by the non-zero suffix bytes, packing two values per header byte to keep the framing tight. The codec is intentionally generic: it works on any byte stream whose elements are width-aligned (1/2/4/8) and have a leading-zero- heavy distribution. It does not require a paired predictor. ## Wire layout ```text [u8 width] element width in bytes (1, 2, 4, or 8) [u32 LE n_elements] element count [packed headers] ceil(n / 2) bytes, each carrying two 4-bit leading-zero-byte counts (high nibble = first elem, low nibble = second elem). [suffix bytes] n_elements groups, each (width - lzcount) bytes; per FPC convention the leading-zero bytes are *implied zero* and never stored. ``` Decoder reverses the layout: for each element read the 4-bit header, then the (width - lzcount) suffix bytes, and zero-extend on the high side. ## Single-predictor variant Burtscher's original FPC pairs two predictors and stores a 1-bit per-element selector. PTWM ships only the single-predictor variant (see [`crate::transforms::PredictorXor`]) and therefore omits the selector bit from the header nibble.