Clean up and consolidate the testing stack

[NSagan271]

Difficulty: 🟢🟡 Good first issue → Intermediate (naturally splits into small PRs)

Scope: Flexible — can be done a slice at a time (one test directory or one stale module per PR), so a newcomer can take a small bite and a regular can take a bigger one.

Subsystems: test/ · CI (.github/workflows/ci.yml) · pyproject.toml

Prerequisites: Basic pytest. Useful for getting familiar with the repo — touching the tests is a good way to learn each subsystem.

Problem

The test tree is a mix of genuinely useful tests and stale/internal testing code, and it's
inconsistent about whether things are actually runnable under pytest:

• Not uniformly pytest. Of ~46 test_*.py files, many are script-style —
  argparse + __main__, meant to be run by hand as python test/.../foo.py
  against a live server (e.g. the per-model client scripts under
  test/orpheus/,
  test/bagel/,
  test/pi05/,
  test/qwen3-omni/). These are useful as manual smoke
  tests / examples but aren't collectable by pytest and aren't clearly labeled
  as manual.
• Stale / debug leftovers. test/scratch/ is debug
  scripts, and some tests elsewhere are out of date with the current code and
  need updating or deleting.
• The real suites are unmarked. The actual automated tests live mostly in
  test/modular/ and
  test/integration/ (the latter has a
  conftest.py), but there's no [tool.pytest.ini_options] in
  pyproject.toml declaring testpaths or markers, so
  there's no canonical "this is the test suite" entry point.
• CI doesn't run tests. The CI workflow
  (ci.yml) only runs ruff check — no test
  job at all, so nothing guards against regressions.

Suggested tasks (each can be its own PR)

• Triage one directory at a time: for each test, decide keep-as-pytest,
  convert to pytest, relabel as a manual script/example, or delete.
• Move or clearly mark the manual client scripts (the per-model
  *_request.py files) so they aren't mistaken for automated tests — e.g.
  relocate under examples/ or a test/manual/ area, or mark with a skip
  so pytest ignores them by default.
• Delete genuinely dead tests (start with test/scratch/).
• Add a [tool.pytest.ini_options] block to
  pyproject.toml with testpaths and markers (e.g.
  gpu, integration, slow) so the fast CPU-only suite is one command.
• (Stretch / separate PR) Add a CI job that runs the CPU-only,
  no-GPU-required subset on PRs, alongside the existing ruff step.

Acceptance criteria

• pytest with no arguments collects and runs a well-defined, green suite (no
  errors from importing manual scripts).
• Manual/demo scripts are clearly separated from automated tests.
• test/scratch and any other confirmed-dead tests are gone.
• (If the stretch goal is included) CI runs the CPU-only test subset on PRs.

Notes

• Coordinate before deleting — confirm a "stale" test is actually obsolete and
  not just temporarily broken. When in doubt, open the PR small and ask in review.
• GPU/model-weight-dependent tests can't run in standard CI; the goal is to make
  the CPU-only subset reliably runnable, not to force everything into CI.

New to M*? Skim How it works and the Contributing guide first.