## Motivation GPTQ / AWQ-style quantized checkpoints store weights with rounded mantissas: an INT4-quantized weight, after dequantization back into BF16 / FP16, has a mantissa whose low bits are always exactly zero. After PTWM's standard `bit_reorder_ieee` + `byte_split` chain, the mantissa plane is a flat byte stream where every element shares the same uniform run of trailing zero bits. Without exploiting this the downstream entropy coder sees a value range that is 2^k times larger than the true alphabet, which costs roughly k bits per element of compression on GGUF / GPTQ-style checkpoints. ## Scope of this op The op handles the dominant case: a *uniform* `k` (1..=7) that holds across every byte in the plane. The chain author (or the explorer) picks `k` at chain-build time and the op stores it as a one-byte parameter on the wire. At encode time the op validates that every byte's low `k` bits are zero; if any byte fails the check the op returns an error and the trial-encode dispatcher falls back to a chain that does not include this op. A non-uniform per-element variant (the "run-length" reading of the issue title) would need a side bitmap and a sparse value buffer; it is deliberately out of scope here. The uniform case captures the published gap and stays a simple invertible byte transform.

`k` is the wire-format parameter (1..=7). `k == 0` is rejected at construction because it would make the op a pure no-op and adds only cost; `k >= 8` is rejected because shifting a `u8` by 8 or more zeroes every byte and loses any high-bit information.