Refactor standalone Windows Nsight discovery into a dedicated module #2609

Motivation

Follow up on PR #2586 with a semantics-neutral internal cleanup proposed in this review comment.

Standalone Windows Nsight discovery currently places product-specific registry and executable-layout knowledge in find_nvidia_binary_utility.py. Extracting that logic would leave the generic finder focused on search policy and orchestration while giving the Windows Nsight details a clear owner.

This refactor should be done after PR #2586 is merged.

Proposed design

Add:

cuda_pathfinder/cuda/pathfinder/_binaries/windows_nsight.py

The module would own the Nsight-specific registry and layout definitions and expose internal lazy candidate iterators:

_REGISTRY_ROOT = ...
_NSYS_TARGET_DIR_BY_ARCH = ...
_NCU_TARGET_DIR_BY_ARCH = ...

def _installed_product_root(product: str) -> str | None: ...
def nsys_candidate_paths() -> Iterator[str]: ...
def ncu_candidate_paths() -> Iterator[str]: ...

The generic finder would continue to own:

• Search orchestration and caching
• Executable validation
• Absolute-path normalization
• Candidate resolution

Its Nsight call sites would become approximately:

from cuda.pathfinder._binaries import windows_nsight

return _resolve_candidate_paths(windows_nsight.nsys_candidate_paths())
return _resolve_candidate_paths(windows_nsight.ncu_candidate_paths())

Lazy candidate iterators avoid a dependency from windows_nsight.py back to the generic finder. They must also preserve the current ncu behavior: yield ncu.bat first and determine the native machine architecture only if the resolver requests the direct executable fallback. A valid launcher should continue to work without invoking architecture detection.

Behavior to preserve

• Keep winreg dynamically imported inside _installed_product_root() so importing cuda.pathfinder remains safe on non-Windows hosts.
• Import windows_nsight itself unconditionally because tests simulate Windows on Linux by patching IS_WINDOWS after import.
• Treat an absent product registration as a normal miss.
• Raise a contextual RuntimeError for incomplete or invalid registry data, and allow access errors to propagate.
• Preserve native-machine selection and the lack of cross-architecture fallback for direct nsys and ncu executables.
• Keep all new names internal, with no public API or observable search-order change.
• Leave Compute Sanitizer in the generic finder for now because its special path is a Windows CUDA Toolkit layout rather than standalone Nsight product discovery.

If more Windows-specific CUDA Toolkit utility layouts are added later, that may justify a separate windows_ctk.py boundary.

Testing

• Move direct registry and Nsight-layout tests to test_windows_nsight.py.
• Keep public-finder, wheel/Conda precedence, terminal-miss, caching, and overall search-order tests in test_find_nvidia_binaries.py.
• Move existing test-authorship markers with their tests unchanged.
• Verify that lazy ncu candidate evaluation does not call native architecture detection when ncu.bat resolves successfully.

Acceptance criteria

• find_nvidia_binary_utility.py owns generic discovery policy and orchestration.
• windows_nsight.py owns standalone Windows Nsight registry and layout knowledge.
• The public API, candidate precedence, caching, error behavior, path normalization, and observable search order remain unchanged.
• Existing tests continue to pass after being reorganized.