Design notes¶
These are the engineering records behind the native engines and the maintainability work. They are kept as written, in their original language, because other documents, tests and the changelog cite them; read them as the reasoning at the time a decision was made, and the API and support matrix as the current contract.
Two numbering schemes
P0–P3 means two different things in this repository. Always check
which scheme a document is using.
| Maintainability refactor | Native-core notes | |
|---|---|---|
| P0 | C ABI contract, cross-language manifest, engine.cu split |
pure-Python baseline |
| P1 | Python backend capability contract | Rust CPU engine |
| P2 | Rust module split | CUDA whole-batch engine |
Native core¶
- Native-core RFC: the Rust and CUDA core, its compatibility contract, engine rules, benchmark protocol and delivery plan.
- P0: baseline: the historical pre-native correctness and performance capture the native engines were compared against; not a description of any release.
- P1: Rust CPU engine
- P2: CUDA whole-batch engine
- P4: CUDA tuning
- Native penalty completion plan:
penalty="l1"on the native CUDA engine (C ABI 2 / Python API 4).
Dispatch and maintainability¶
- CPU auto-dispatch RFC: when
backend="auto"may select the Rust CPU engine, from bounded runtime evidence. - Maintainability refactor (中文): what the P0–P3 audit phases changed and why.
Repository-level working rules, including the invariants that break silently,
are in
AGENTS.md.