# Changelog

## Score

- CAI 64 → 65 (+0.9)

## Lenses

- Code Health 68 → 68 (+0.0)
- Architecture 98 → 98 (-0.0)
- Maturity 61 → 65 (+3.5)
- Readiness 64 → 64 (+0.0)
- Security 62 → 62 (-0.3)

## Resolved (7)

- Inconsistent destructor naming. While Rust typically uses `Drop` trait, this API exposes explicit `drop` methods on several types (`CublasContext`, `Event`, `ArrayObject`, `Context`). However, `CudnnContext` does not have a `drop` method listed, nor does `CudaBuilder`. This suggests an inconsistent resource management strategy across the library's context and handle types.
- Inconsistent return type for getter methods. `get_atomics_mode` and `get_math_mode` return `Result` (likely `Result<T, CublasError>`), but the input parameter for `set_atomics_mode` is a `bool` while `set_math_mode` takes a `MathMode` enum. While the setter signatures are distinct due to parameter types, the getter return types being `Result` without a clear error case (getting a mode rarely fails) is suspicious. More importantly, `CudnnContext` does not expose similar mode getters/setters in the same way, making cross-library consistency hard to judge, but within `CublasContext`, the `Result` return for simple getters is inconsistent with typical Rust patterns where getters return `T` or `Option<T>` unless they can fail.
- Inconsistent version reporting API. `CublasContext` and `CudnnContext` return a tuple `(u32, u32, u32)` for their respective library versions. `CudnnContext` also has a separate `cuda_version()` method returning the same tuple structure. In contrast, `CudaApiVersion` provides a dedicated type with `get()`, `major()`, and `minor()` methods. This mixes tuple-based returns with dedicated accessor types for similar information.
- Most significant orphaned file (crates/rustc_codegen_nvvm/src/debug_info/metadata/enums.rs)
- Naming inconsistency between strided and non-strided variants. The API exposes pairs of methods where one is suffixed with `_strided` and the other is not (e.g., `amin` vs `amin_strided`). However, the non-strided versions implicitly assume a stride of 1, while the strided versions take explicit stride parameters. This is a common pattern, but the naming is inconsistent with the `cuda_std` module which uses `WarpShuffleValue` methods that do not have a 'stride' suffix but rather specific modes. More critically, within the BLAS level 1 operations, some operations like `rotg` and `rotmg` do not have a non-strided counterpart in the list, while others like `amin` do. This creates an inconsistent surface where users must check if a non-strided version exists for every operation.
- Off-boarding risk: anonymized user #1
- PR-triggered workflow without a permissions block

## New (119)

- Medium vulnerability: [GHSA redacted] (Cargo.lock)
- Most significant orphaned file (crates/optix/src/acceleration.rs)
- Off-boarding risk: anonymized user #1
- Orphaned file (crates/blastoff/src/context.rs)
- Orphaned file (crates/blastoff/src/error.rs)
- Orphaned file (crates/blastoff/src/level1.rs)
- Orphaned file (crates/blastoff/src/level3.rs)
- Orphaned file (crates/blastoff/src/lib.rs)
- Orphaned file (crates/blastoff/src/raw/level1.rs)
- Orphaned file (crates/blastoff/src/raw/level3.rs)
- Orphaned file (crates/cuda_std/src/atomic/mid.rs)
- Orphaned file (crates/cuda_std/src/float.rs)
- Orphaned file (crates/cuda_std/src/float_ext.rs)
- Orphaned file (crates/cuda_std/src/io.rs)
- Orphaned file (crates/cuda_std/src/ptr.rs)
- Orphaned file (crates/cuda_std/src/rt/driver_types_sys.rs)
- Orphaned file (crates/cuda_std/src/rt/error.rs)
- Orphaned file (crates/cuda_std/src/rt/mod.rs)
- Orphaned file (crates/cuda_std_macros/src/lib.rs)
- Orphaned file (crates/cudnn/src/activation/mod.rs)
- …and 99 more
