cargo oxide doctor cannot diagnose missing CUDA headers because cuda-bindings panics first #87

Summary

On a machine without CUDA headers installed, cargo oxide doctor cannot reach the doctor checks. The cargo alias first builds cargo-oxide, which pulls cuda-bindings; that build script falls through to bindgen and panics when cuda.h is missing.

Repro

From a fresh clone on a host without /usr/local/cuda/include/cuda.h:

cargo oxide doctor

Observed failure:

warning: cuda-bindings@0.1.0: cuda-bindings: Could not read cuda.h at /usr/local/cuda/include/cuda.h: No such file or directory
wrapper.h:1:10: fatal error: 'cuda.h' file not found
thread 'main' panicked at crates/cuda-bindings/build.rs:70:10:
Unable to generate CUDA bindings: ClangDiagnostic(...)

Expected

The build script should report missing CUDA headers directly and avoid a bindgen panic. Related: CUDA redistributable/cross layouts may place headers under target-specific directories such as targets/x86_64-linux/include or targets/sbsa-linux/include; the build script should not hardcode a single target directory.

Notes

PR #88 updates the build script to:

• discover cuda.h in both include/ and targets/*/include/
• use the discovered header for version/cfg probing
• pass all discovered include directories to bindgen
• add matching target-layout library paths for discovered target roots
• return a direct build-script error for missing/unreadable headers instead of panicking in bindgen