Testing And Quality Assurance

Testing & Quality Assurance

Running the suite

From repo root with .venv active:

.venv/bin/python -m pytest                    # full suite
.venv/bin/python -m pytest -m unit            # unit only
.venv/bin/python -m pytest -m "not slow"      # skip slow ML tests
.venv/bin/python -m pytest tests/test_basic.py -v

pytest.ini config:

• testpaths = tests
• --strict-markers — unknown markers fail.
• Filters out transformers/torch deprecation noise.

Markers (enforced)

Always mark a test. Unmarked tests run in every job and slow CI down.

Fixtures (from tests/conftest.py)

The mock is wired by patch("src.database.dictionary_db.DictionaryDB", return_value=mock_db) — i.e. patching the class before app is imported. Do not later try patch("app.dict_db", ...); the singleton is already the mock and re-patching at the module level fights the class-level patch.

⚠️ The auth_headers fixture is broken

It sets sess["authenticated"] and sess["user"], but the production gate checks sess["logged_in"] + sess["user_email"] + sess["session_id"] (app.py). Until the fixture is fixed, build the session inline in your test:

def test_authenticated_route(client, flask_app):
    with client.session_transaction() as sess:
        sess["logged_in"] = True
        sess["user_email"] = "test@example.com"
        sess["user_role"] = "admin"
        sess["session_id"] = "test-session"
    # Also bypass the single-session DB check
    with patch("app.get_user_db") as gud:
        gud.return_value.is_session_valid.return_value = True
        resp = client.get("/api/dictionary/search?q=ran")
    assert resp.status_code == 200

If you fix auth_headers, also patch is_session_valid inside it.

Configuring mock_db per test

def test_search(client, mock_db):
    mock_db.search_words.return_value = [
        {"chuukese_word": "ran", "english_translation": "water"}
    ]
    # OR via raw collection:
    mock_db.dictionary_collection.find.return_value = iter([
        {"chuukese_word": "ran", "english_translation": "water"}
    ])

The methods that exist on the real DictionaryDB (and that you should mock):

• search_word, search_words, add_word
• search_phrases, add_phrase

There is no search_entries, bulk_insert_entries, etc. Don't mock methods that don't exist — the real code will never call them.

Translation tests

@pytest.mark.translation
@pytest.mark.slow
def test_helsinki_chk_to_en():
    from src.translation.helsinki_translator_v2 import HelsinkiTranslator
    t = HelsinkiTranslator()
    t.setup_models()  # no direction arg
    assert t.translate("ran", "chk_to_en") == "water"

Use BLEU only when you control the reference set — small samples are noisy.

Test file inventory

The tests/ directory mixes real pytest tests and exploratory scripts. Files prefixed test_ are collected:

• test_basic.py — publication manager, jworg lookup. Some assertions hit the network — mark integration.
• test_collections.py — DB collection ops.
• test_helsinki_trainer.py — fine-tuner helpers (mark slow).
• test_translation.py — end-to-end translate (mark translation/slow).
• test_scripture_parsing.py — pure unit (regex, book lookup).
• test_word_families.py — pure unit.

Scripts without the test_ prefix (e.g. find_complex_words.py, debug_models.py) are not test runners — they exist for one-off investigation.

CI guidance

The default CI invocation should be pytest -m "not slow and not translation and not integration" to keep runs fast. Translation/integration jobs run on a separate, slower lane.

Pitfalls

• Importing app triggers DB init — that's why fixtures patch DictionaryDB before the import. Don't import app at module top of a test file; let the fixture do it.
• The _safe_open in conftest blocks file opens outside config/ and models/ — if your test needs to read a file, point it under config/ or extend the allowlist.
• Two gunicorn workers in prod ≠ test fixture state. Don't write tests that assume cross-request in-memory state.
• WTF_CSRF_ENABLED=False is set in the fixture, but the app does not actually use Flask-WTF — the flag is a no-op kept for legacy reasons.