[tpn]

Superseded by #12088, which now owns this implementation within its complete feature family. The original commits, branch and validation evidence remain available. This draft is closed unmerged.

---

Historical scope and validation below refer to the original source heads.

coop::WarpTranspose and coop::WarpTransposeTimesliced add CUB warp-based exchanges to block Load/Store. Both require full-block participation and a thread count divisible by 32. Store permutes values and validity bits into warp-striped order, including zero-count writes; the timesliced algorithm uses a smaller exchange buffer.

Review after #12106. This draft owns the selectors, native mappings, scratch queries, tests, example and documentation: 14-file owning diff. GitHub’s main-based Files changed view includes prerequisites. It uses the migrated, rustc-free Oxide analyses through NVIDIA/cuda-rust#1593.

Current-source host/docs and provider MSRV checks, 798 native requests, both original DEV consumers and all 36 compiler controls pass. GPU validation passes 2,906 combined cases, the 11,436-value example, four sanitizer runs and three expected device failures, each followed by a successful health check. Device inspection confirms launch geometry, scratch and typed validity ABI. The minimal example has no stack/local traffic; the observation fixture retains helper stack/local traffic. No timing claim is made.

Rust host CI, pre-commit.ci and the project check pass on this head.