test: share one helper for isolated-subprocess test snippets

[aryanputta]

Description

closes #2412

Five tests run a self-contained snippet in a fresh interpreter with subprocess.run([sys.executable, "-c", code], ...), each to keep a process-global side effect from reaching the rest of the suite: the CPython pending-call queue, the import-time rlcompleter patch, and a deliberate glibc abort. They had drifted apart on which of the surrounding hazards they handled.

The sharpest one is the working directory. python -c puts the parent's cwd at the head of the child's sys.path, so a run started from cuda_core/ makes the child import the cuda/core/ source tree instead of the installed wheel. Two sites work around it and carry a comment explaining it; the trap is rediscovered per site.

cuda_python_test_helpers.subprocess_runner.run_python_snippet is now the single place these are handled:

run_python_snippet(code, *, cwd, timeout=None, extra_env=None, unset_env=(), check=True)

cwd is required rather than defaulted, so the shadowing hazard cannot be forgotten. Output is always captured as text so a failure message can quote both streams. stdin is always closed, because PYTHONINSPECT keeps the interpreter alive after -c and an inherited stdin leaves the implicit REPL waiting.

All five call sites move to it:

• cuda_core/tests/test_rlcompleter_patch.py (_run_probe and the opt-out probe)
• cuda_core/tests/graph/test_graph_definition_lifetime.py (two sites)
• cuda_bindings/tests/test_cuda.py

This is a pure refactor. unset_env is opt-in per call site, so the two graph tests and the bindings test keep inheriting PYTHONPATH exactly as they do today, and only test_rlcompleter_patch.py drops it as it already did. _run_probe keeps check=False because both of its callers assert on the exit code themselves. unset_env is applied before extra_env, which preserves _run_probe's clear-then-set handling of CUDA_CORE_DONT_FIX_TAB_COMPLETION; a test pins that ordering.

Net across the five sites is 21 added against 59 removed.

The issue counted four sites. The fifth, test_graph_definition_lifetime.py:790, arrived with #2406 on the same day the issue was filed.

The description sketched a run_in_subprocess fixture that re-runs the current test node with pytest <nodeid>. None of the five sites have that shape, and re-running the node would inherit pytest's rootdir, which is the exact cwd hazard two of them work around. Rationale is in #2412.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[aryanputta]

Pushed 872ddc2 on top, a self-review pass against @rwgk's guidance on #2196 and #2239 rather than anything new in the behavior.

• Cut the module docstring down to one line. The version I opened with narrated the refactor and named the specific side effects in other test files. That is PR-description material, and it drifts.
• Removed the comments that restated the code below them or pointed at _run_probe by name from a test in a different file.
• Kept one general statement of the cwd contract, on the argument itself, since that is what makes the argument required.
• Added assert cwd instead of letting an empty string reach subprocess.run, in the shape suggested for _resolve_in_trusted_dirs on feat(pathfinder): add CTK-root canary fallback for NVIDIA binaries #2196, plus a test that the guard fires.

Diff is now 27 added against 50 removed on those files. No behavior change from the first commit.

[mdboom]

As you pointed out, this more correctly belongs in cuda_python_test_helpers, but we don't currently have tests there.

It feels like over-testing to me to have these tests -- we are already exercising the required behavior in the other tests, so I say just delete this file.

[mdboom]

I think we could make this optional. When not provided, use an auto-cleaned-up tempdir for cwd. That should simplify most (but not all) of the call sites.

[aryanputta]

@mdboom The requested changes are addressed in 3286d4396b: removed cuda_core/tests/test_subprocess_runner.py, made cwd optional with an auto-cleaned temporary directory, and simplified the call sites. The old review threads are now outdated. Could you please re-review this head when convenient?

[mdboom]

LGTM. Just needs the merge conflicts resolved and another round of tests and I think this is good-to-go.