test: convert legacy GPU diagnostic scripts into maintained pytest coverage

[TheHiddenObserver]

Context

PR #79 completed its maintained physical-GPU acceptance suite on a Tesla P100 with 33 passed, 2 skipped, 0 failed. During the same remote run, several ignored or script-style files under dev/tests/ were also executed manually. They are not part of the maintained pytest Gate and currently produce collection/runtime failures or zero collected tests.

This issue tracks cleanup and promotion of any still-useful checks into maintained, deterministic pytest coverage. It is a follow-up and does not block PR #79.

Affected legacy scripts

• dev/tests/test_coxph_3backends.py
  • executes work at module import time;
  • currently reaches a CuPy CoxPH runtime error on the P100 environment.
• dev/tests/test_irls_gpu.py
  • requires a missing loss_name fixture;
  • currently ignored by .gitignore.
• dev/tests/test_lasso_cv_torch_quick.py
  • script-style execution at import time;
  • pytest collects zero tests.
• dev/tests/test_ridge_cv_torch_backend.py
  • script-style execution at import time;
  • pytest collects zero tests.
• dev/tests/test_torch_comprehensive.py
  • script-style execution at import time;
  • pytest collects zero tests.
• dev/tests/test_lassocv_inference_simple.py
  • uses an older inference API contract and is rejected by current validation.

Goals

1. Decide for each file whether to:
   • convert it into maintained pytest tests;
   • replace it with existing PR79/backend regression coverage;
   • move it to an explicitly named manual/diagnostic directory; or
   • delete it as obsolete.
2. Remove all module-import side effects from tests.
3. Add explicit CUDA/CuPy/Torch capability markers and deterministic skip reasons.
4. Eliminate undeclared fixture dependencies.
5. Align inference tests with the current public API and strict unsupported-combination contracts.
6. Ensure maintained GPU tests run from a clean checkout rather than relying on ignored local files.

Required base/head triage

For runtime failures that exercise production code—especially CoxPH and LassoCV inference—run the minimized reproducer on both:

• PR79 base: a4879fb4d9fb183efc01f147cd2cc501691f28c4
• current post-PR79 code

Classify each result as:

• verified pre-existing production issue;
• obsolete test expectation;
• test harness defect; or
• new regression.

Do not label a production-path failure as pre-existing without this base/head comparison.

Acceptance criteria

• No retained test performs substantial computation at import time.
• Every retained file contains discoverable pytest test functions/classes.
• GPU requirements use explicit markers/skip conditions.
• Missing fixtures are defined or the dependency is removed.
• Current inference contracts are asserted explicitly.
• Base/head classification is recorded for production-path failures.
• Maintained CPU test suite remains green.
• Maintained physical-GPU suite remains green.
• .gitignore rules are simplified so maintained tests are tracked and manual scripts are clearly separated.

Suggested structure

dev/tests/
  test_*_gpu.py              # maintained pytest coverage

dev/manual/
  gpu_diagnostics/           # intentionally manual scripts

Environment observed during PR79 follow-up

• Python 3.9.16
• CuPy 13.6.0
• Torch 2.0.0+cu117
• Tesla P100-SXM2-16GB
• pytest 8.4.2

Related: PR #79.

[TheHiddenObserver]

The maintained-test cleanup and governance work tracked by this issue was completed in PR #87 and is included in the 0.2.4 release tree.

Completed work:

• established dev/tests/test_*.py as tracked, discoverable, side-effect-free maintained pytest coverage;
• simplified .gitignore so maintained tests cannot remain silently ignored;
• established dev/manual/gpu_diagnostics/ for intentionally manual or hardware-specific scripts;
• documented explicit GPU capability checks, deterministic skip reasons, current inference contracts, clean-checkout requirements, and exact-head hardware evidence rules;
• triaged every historical filename listed in this issue as retired, replaced by maintained coverage, obsolete, or a test-harness defect;
• added maintained Torch/CuPy and estimator/backend regressions;
• completed the full CPU and exact-head physical-GPU acceptance suites.

One historical item cannot be completed literally: test_coxph_3backends.py was an ignored/unversioned local diagnostic and its original source is unavailable from Git history. The historical CuPy CoxPH failure therefore cannot honestly be reproduced on the PR #79 base and current head. The repository documentation explicitly records that no claim is made that this exact historical failure was pre-existing or fixed; a recovered reproducer would need a new base/head investigation.

Current maintained Cox backend tests and physical-GPU acceptance are green, so there is no presently reproducible production defect to keep this broad cleanup issue open for.

Implementation merge: PR #87, commit 249456610c9d8ab62e9f2749e1fc7a769250b08a.
Release merge: PR #88, commit 0aeeb95b60e3e274053b8f1b6427ae50c8eec015.

Closing as not planned rather than completed because the unavailable historical reproducer prevents literal completion of that single base/head classification checkbox. Any recovered or newly reproducible CuPy CoxPH defect should be filed as a focused new issue with a minimized reproducer.