[FIX]: pathfinder: Add missing static-lib CUDA 12 path for Windows conda layouts.

[rwgk]

Latent bug discovered while testing PR #1977 (xref issue #1038) with Windows Conda.

Source change — find_static_lib.py

• The _StaticLibInfo schema had a single conda lookup directory per library (conda_rel_path: str). On Windows that was hard-coded to Library/lib/x64, but only CUDA 13 conda environments lay cudadevrt.lib out that way; CUDA 12 conda environments put it in Library/lib. So on a Windows + CUDA 12 conda setup, pathfinder skipped the conda anchor entirely and fell through to CUDA_PATH, often producing a misleading "not found" error pointing at the wrong location.
• The fix changes the schema to a tuple of lookup paths (conda_rel_paths: tuple[str, ...]) so each lib can declare every layout it knows about. For cudadevrt:
  • Windows: ("lib/x64", "lib") — try the CUDA 13 layout first, fall back to the CUDA 12 layout.
  • Linux: ("lib",) — unchanged, single entry.
• _FindStaticLib.try_with_conda_prefix now loops through self.conda_rel_paths and returns the first hit instead of probing one fixed subdirectory.

Test change — test_find_static_lib.py

• Existing search-order test (test_locate_static_lib_search_order) is updated mechanically: CUDADEVRT_INFO["conda_rel_path"] → CUDADEVRT_INFO["conda_rel_paths"][0].
• New test test_locate_static_lib_conda_rel_path_fallback:
  • Asserts the fallback semantic by injecting a deliberately-missing first entry ("missing-first") when the platform only ships a single path. On Windows, where the tuple already has two entries, it uses them as-is.
  • Drops the file at the second listed path under the conda anchor and verifies locate_static_lib("cudadevrt") finds it with found_via="conda".
  • Real coverage: confirms the loop tries entries in order and a missing first entry doesn't abort the search.

Real (not mock-based) testing was reported here:

• WIP: pathfinder_compatibility_guard_rails #1977 (comment) — PR 1977 conda testing results analysis

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rwgk]

/ok to test

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2015/
https://nvidia.github.io/cuda-python/pr-preview/pr-2015/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2015/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2015/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.