build: use the validated NVCC compiler

[morluto]

Fixes #1.

Problem

setup.py validated the compiler selected by NVCC, but the Make invocation used MOK_NVCC (falling back to nvcc). An NVCC-only configuration could therefore validate one CUDA toolkit and compile with another compiler resolved from PATH.

Change

• Select MOK_NVCC, then NVCC, then nvcc as the compiler command.
• Return the resolved executable from check_nvcc().
• Pass that exact resolved executable to Make.
• Document the compiler override in the installation instructions.

This preserves MOK_NVCC support while making validation and compilation use one compiler path.

Regression coverage

A focused harness loads setup.py, exercises both environment-variable paths, stubs a CUDA 13 NVCC version response, and captures the Make command. It verifies that an NVCC-only selection passes the validated executable path to Make rather than NVCC=nvcc.

Validation

• python -m py_compile setup.py
• git diff --check
• Focused NVCC selection/build-command regression harness — passed

---

Note

Low Risk
Build/install wiring only; no runtime kernel or training behavior changes.

Overview
Fixes a mismatch where setup.py could validate one compiler (via NVCC) but invoke Make with another (MOK_NVCC or bare nvcc on PATH).

Compiler selection is now centralized: MOK_NVCC → NVCC → nvcc, resolved with shutil.which, version-checked in check_nvcc(), and that same resolved path is passed to make as NVCC=.... The README documents overriding the compiler with MOK_NVCC or NVCC for both the check and the build.

^{Reviewed by Cursor Bugbot for commit 25e607c. Bugbot is set up for automated code reviews on this repo. Configure here.}