Unlike `XorDelta` and `FloatDelta`, which take a *reference tensor* as a second input and produce a cross-tensor residual, `IntDelta` is a single-input intra-plane transform: each element becomes the integer difference from its predecessor in the same plane. For smoothly-varying data (think attention-weight rows whose neighbouring values are numerically close), the integer delta is concentrated near zero and compresses harder under the downstream entropy coder than the raw byte stream does. This is the same preprocessing used by ndzip and FPC for scientific floating-point streams; see [ndzip-gpu (ICS 2021)](https://dps.uibk.ac.at/~fabian/publications/2021-ndzip-gpu-efficient-lossless-compression-of-scientific-floating-point-data-on-gpus.pdf). ## Element-width awareness The delta is computed at the plane descriptor's `element_width`: - `Byte` (1 B) — byte-level wrapping subtraction. - `Word2` (2 B) — pairs interpreted as little-endian u16, wrapping sub. - `Word4` (4 B) — u32 LE, wrapping sub. - `Word8` (8 B) — u64 LE, wrapping sub. - `Nibble` — rejected; semantics are ambiguous for half-byte elements. Round-trip is exact for every input the encoder accepts. The op is pure-byte (no `lossy` feature gate) because integer wrapping is involutive in the same way XOR is.

`forward`: `out[0] = in[0]`, `out[i] = in[i].wrapping_sub(in[i-1])`. `inverse`: `in[0] = out[0]`, `in[i] = out[i].wrapping_add(in[i-1])`. The stride is the plane's `element_width`. Wrapping arithmetic at that width is exactly invertible.