Cross-checkpoint compression stores a tensor as the *difference* from a reference tensor (e.g. a base model checkpoint). When two checkpoints are close, the residual concentrates on very few bits and compresses far better than the raw tensor. Two modes are implemented here: - **Byte-level XOR** ([`xor_encode`] / [`xor_decode`]): bit-exact and safe for lossless compression of any dtype. - **Float-arithmetic subtraction** ([`float_sub_encode`] / [`float_sub_decode`]): interprets buffers as little-endian `f32` arrays and computes element-wise `raw - reference` / `reference + residual`. Best suited for fine-tune checkpoints where weight perturbations are small. Learned-basis and predictive modes remain out of scope for this module.

# Caveats - **NaN bit patterns are not preserved**: IEEE 754 arithmetic may produce a different NaN payload than the input, so inputs containing NaN values will not round-trip bit-exactly. - **Catastrophic cancellation**: if the residual and reference have extreme magnitude differences the addition loses precision and the codec is **not lossless** for those elements. For trained neural network fine-tune checkpoints where fine-tune deltas are small perturbations of the base model, this is not expected to occur; callers requiring guaranteed bit-exact roundtrip should verify with their data.

Both inputs are interpreted as little-endian `f32` slices. Byte lengths must be equal and divisible by 4. The inverse is [`float_sub_decode`]. # Caveats - **NaN bit patterns are not preserved**: IEEE 754 arithmetic may produce a different NaN payload than the input, so inputs containing NaN values will not round-trip bit-exactly. - **Catastrophic cancellation**: if the raw value and reference have extreme magnitude differences (e.g. `raw = 1.0`, `reference = 2e30`), the subtraction loses precision and the codec is **not lossless** for those elements. For trained neural network fine-tune deltas, where weights are small perturbations of the base, this is not expected to occur; callers requiring guaranteed bit-exact roundtrip should verify with their data.

For equal-length inputs, returns `raw ^ reference` byte-by-byte. XOR is self-inverse, so [`xor_decode`] is identical in effect but named for clarity at call sites.