[nihalpasham]

Fixes #87. PR #172 covered the cuda-bindings half (a clean build-script error instead of a bindgen panic when cuda.h is missing); this PR covers the reachability half: on a machine with no CUDA toolkit and no driver, cargo-oxide now builds, loads, and cargo oxide doctor prints an actionable diagnosis instead of dying in a dependency's build script.

Before: cargo oxide doctor -> cargo builds cuda-bindings -> hard error on missing cuda.h
After:  cargo oxide doctor -> runs, reports each missing piece and how to fix it

What changed

• cargo-oxide drops its cuda-host / cuda-core dependencies. Its graph now contains no cuda-bindings, bindgen, or clang-sys, and the binary has no libcuda.so.1 in DT_NEEDED, so it builds and runs on a bare machine. libnvvm-sys / nvjitlink-sys stay: they are pure-Rust runtime-dlopen crates with zero build-time CUDA cost.
• Run-mode GPU arch auto-detect now shells out to nvidia-smi --query-gpu=compute_cap --format=csv,noheader (the derivation smoketest.sh already uses) instead of the CUDA driver API. Output is trusted only on exit 0 plus a <digits>.<digits> first line, because nvidia-smi prints failure banners to stdout. Precedence is unchanged: --arch and CUDA_OXIDE_TARGET still override, the result stays the advisory CUDA_OXIDE_DEVICE_ARCH hint, and sm_XYa is still emitted for cc >= 9.0.
• libdevice discovery is consolidated into libnvvm-sys as find_libdevice() (libdevice ships in the toolkit's nvvm/ component next to libnvvm.so), removing a duplicated cuda_roots_from_env. cuda_host::ltoir::find_libdevice remains as a thin wrapper with identical signature and error text.
• doctor is build-free: a new side-effect-free resolve_doctor_context() (with backend::backend_so_candidate()) replaces resolve_context(), which used to build the backend unconditionally, or even git-clone in standalone mode, before a single check ran.
• doctor gains two checks:
  • CUDA headers (cuda.h), fatal: mirrors cuda-bindings/build.rs discovery (CUDA_TOOLKIT_PATH > CUDA_HOME > /usr/local/cuda; include/ then targets/<x86_64-linux|sbsa-linux>/include per host arch). On failure it lists every probed path and the env vars to set. The probe carries a lockstep comment: build.rs logic is not importable and cuda-bindings must stay out of the graph, so the two probes are kept in sync by hand. Both helpers are pure and unit-tested.
  • NVIDIA driver / GPU, informational: reports the GPU name and compute capability via nvidia-smi, with a /proc/driver/nvidia/version fallback that distinguishes "driver loaded but nvidia-smi broken" from "no driver at all". Non-fatal because only cargo oxide run needs a driver; GPU-less CI and cross-compiling are supported workflows.
• The backend .so check downgrades to informational ("not built yet, run cargo oxide setup"): run/build/pipeline build it on demand, so a fresh healthy clone stays green.
• CI: the cargo-oxide unit-test job drops the toolkit install and stub shadow, and a new step runs the built binary's doctor on that toolkit-less runner and asserts the cuda.h diagnosis is reached (an end-to-end regression gate for the exact issue cargo oxide doctor cannot diagnose missing CUDA headers because cuda-bindings panics first #87 environment). examples-compile drops the libcuda stub shadow that existed only for cargo-oxide, in its own commit so it reverts cleanly.

For users with a toolkit and GPU

• run/build/pipeline/debug flows are untouched (they always shelled out to cargo; GPU work happens in the example's own process, which still links cuda-core as before).
• Identical arch values: nvidia-smi's compute_cap is the same cc pair as cuDeviceGetAttribute.
• doctor: same verdicts as before, plus the headers and driver lines.

Known tradeoffs (documented in code)

• nvidia-smi lists GPUs in PCI order while CUDA defaults to fastest-first, so on heterogeneous multi-GPU boxes the hint may describe a different GPU than CUDA device 0. Contained: the hint is advisory (the backend ignores incompatible hints) and --arch / CUDA_OXIDE_TARGET are hard overrides.
• Driver present but nvidia-smi absent or broken: the hint silently becomes None and the backend default applies; doctor's /proc fallback at least diagnoses that state.
• doctor no longer auto-builds the backend; anyone who relied on that side effect now gets an informational "run cargo oxide setup" line.

Verification

• cargo tree -p cargo-oxide -e normal,build: no cuda-bindings / cuda-core / cuda-host / bindgen / clang-sys (all five present before).
• readelf -d cargo-oxide: libcuda.so.1 gone from DT_NEEDED (present before; it is why both CI workflows carried the stub shadow).
• CUDA_TOOLKIT_PATH=/nonexistent CUDA_HOME=/nonexistent cargo build -p cargo-oxide succeeds: no build script in the graph consults those vars anymore.
• doctor on a toolkit machine passes every pre-existing check and shows the two new lines; with CUDA_TOOLKIT_PATH=/nonexistent it prints the probed cuda.h paths and exits 1.
• cargo oxide build vecadd completes clean with a fresh vecadd.ptx and no swallowed-error signatures.
• fmt / clippy -D warnings / doc -D warnings / tests green on cargo-oxide, libnvvm-sys, cuda-host; the three pre-existing detect_run_target_arch tests pass unmodified.

GPU-box confirmation requested: cargo oxide doctor should print the GPU name and compute cap on the driver line, and cargo oxide run vecadd should report a "Detected GPU arch" matching nvidia-smi --query-gpu=compute_cap and pass.

Suggested labels: QoL-improvement, cargo-oxide, host-apis, build-related.

GPU-box confirmation (done)

Verified on a 2x RTX 5090 box: doctor prints "NVIDIA driver / GPU... ✓ NVIDIA GeForce RTX 5090 (compute capability 12.0)" and exits green; cargo oxide run vecadd reports "Detected GPU arch: sm_120a (via nvidia-smi)" and passes with all 1024 elements correct. readelf -d on the built binary confirms libcuda.so.1 is gone from DT_NEEDED.