[cschanhniem]

Description

CI's pytest -q step has been failing at collection with ModuleNotFoundError: No module named 'scripts', which aborts the entire run — no tests execute on any PR or on main. This has masked real regressions. Two root causes, both fixed here.

1. Test collection (pyproject). [tool.pytest.ini_options].pythonpath was ["src", "scripts"] but did not include the repo root. The scripts directory is a real package (scripts/__init__.py) that 26 test modules import as a top-level package (import scripts.<mod>). With only src and scripts on the path, import scripts is unresolvable. It happened to work under python -m pytest (which adds the CWD to sys.path) but not under the pytest console script that CI runs. Fix: add "." so the repo root is on the path under both invocations.

2. CLI subparser regression (src/openamp_foundry/cli/main.py). With collection fixed, 148 tests failed on NameError: name 'subparsers' is not defined from the CLI. The experiment-priority-check subparser was registered via subparsers.add_parser(...), but the subparsers object is named sub (92 other call sites use sub; this was the lone subparsers). A regression from #832 that CI never caught because collection was already broken. One-identifier fix.

Change class

• Class B — infrastructure / CLI change

Type of change

• Bug fix
• Infrastructure / CI / tooling

Safety impact

• No biological misuse capability added
• No harmful optimization objective added
• No efficacy, safety, clinical, or therapeutic claims
• No scoring, ranking, calibration, benchmark-threshold, or safety-policy behavior changed
• Limitations documented (see below)

Verification

Run with the CI-equivalent invocation (pytest -q, console script) after pip install -e ".[dev]":

• Before: collection aborts — 3 errors during collection, 0 tests executed.
• After: 4418 passed, 8 skipped.
• ruff check src/openamp_foundry/cli/main.py → clean.
• The three originally-uncollectable benchmark modules now collect and pass (14 tests).

Limitations — pre-existing failures this unmasks

Fixing collection reveals 4 failing tests (2 distinct pre-existing bugs) that fail on unmodified main, are unrelated to this change, and were hidden by the broken collection. They are out of scope here because both require product-level judgment and one is an explicit Class-C stop condition (scoring). Filed as separate issues:

1. tests/evidence/test_candidate_summary_card.py::test_lowercase_sequence_fails — the validator .upper()-normalizes before checking amino acids, so it accepts lowercase, but its own test demands rejection (and presynth_check.py also normalizes). Code-vs-test conflict needing a spec decision.
2. tests/evidence/test_negative_result_informativeness.py::TestCLI (×3) — the informativeness scorer classifies its own test fixture below the INFORMATIVE threshold (≥6.0/7.0). Changing scoring behavior is a Class-C / human-review item.

This PR is deliberately the minimal, verified fix for the named bottleneck (CI cannot run the test suite) — one loop, one change. It is a strict improvement: from 0 tests executing to 4418 passing.

Agent disclosure

• Fully or partly agent-generated and human-reviewed

Reviewer focus

• Confirm adding "." to pythonpath is acceptable (repo root on sys.path; no test relies on the bare benchmarks. top-level — all 26 use scripts.).
• The two pre-existing bugs above need owners; they are not addressed here on purpose.

🤖 Generated with Claude Code

---

Generated by Claude Code