Review Topic Test Hygiene

Review topic: Test hygiene & coverage of changed behavior

When this applies

Load and apply this skill whenever the diff shows any of:

• New/modified runtime code (block, model I/O contract, endpoint, SDK/CLI path, execution-engine rule) with no new/changed test, or only a test that mocks out the very thing that changed.
• A new/edited dispatch on a runtime selector: enum→lambda maps (e.g. PROPERTIES_EXTRACTORS), if kind == ..., block type/property_name switches, get_execution_engine_compatibility.
• A new @pytest.mark.skipif/skip/slow or env-gated test (os.getenv("SKIP_*_TESTS", True), RUN_*).
• New global test state: module-level monkeypatch, sys.modules[...] = stub, os.environ[...] = at import time, or autouse fixtures mutating shared singletons/registries.
• A GPU/hardware-only test whose CI runner is unclear.

Review checklist

Tag each finding BLOCK / FLAG / NIT. See Standards below for the one canonical statement of each rule.

• BLOCK — Changed runtime behavior ships with no test that actually runs in the PR's x86 CI (Standard 1), OR a bugfix lands with no regression test reproducing the bug.
• BLOCK — A new env-gated / GPU test's gate is never flipped False by any CI job → dead test giving false confidence (Standard 4).
• BLOCK — A new inference_models/tests/unit_tests gpu_only test lacks a dependency marker selected by .github/workflows/unit_tests_inference_experimental_gpu.yml, so no GPU job collects it (Standard 4).
• BLOCK — A leaked module-level monkeypatch, sys.modules stub, or mutated global registry corrupts other tests (Standard 3).
• FLAG — A new selector/branch/enum value is exercised only by literals, not by a test driving that runtime value (Standard 2).
• FLAG — Mock-only test that asserts the mock's own return value rather than the changed output (Standard 6).
• FLAG — New marker used in the diff but not declared in pytest.ini → silent no-op deselection (Standard 5).
• NIT — Brittle exact-value assertion on a numerically unstable op (Standard 7).
• NIT — Broken/stale test asset, stub, or a test not updated alongside the API it covers.

Not blocking

• Do NOT demand a test that duplicates an existing behavior test at a lower layer (e.g. a unit test when the same path is already driven end-to-end by an integration test that runs in CI).
• Do NOT demand exact-float assertions on model predictions — shape/contract/tolerance is the correct bar; over-tight asserts get muted and erode coverage.
• Do NOT block a GPU/hardware test for running only on the self-hosted runner if the same behavior also has an x86-CI test; the GPU test is additive.
• Pure refactors with no behavior change need no new test, only that existing tests still pass and were updated to the new signature.

Standards

1. Changed behavior is verified by a test that runs in the PR's CI. Map each behavior in the diff to the covering test AND the CI workflow that executes it:

   • unit workflows → unit_tests_workflows_x86.yml (tests/workflows/unit_tests)
   • workflow execution → integration_tests_workflows_x86.yml
   • inference_models I/O → integration_tests_inference_models.yml (tests/inference/models_predictions_tests)
   • HTTP server → integration_tests_inference_server_x86.yml (tests/inference/integration_tests)
   • SDK → unit_tests_inference_sdk_x86.yml (tests/inference_sdk/unit_tests)
   • CLI → unit_tests_inference_cli_x86.yml (tests/inference_cli/unit_tests)

   A test that is skipped-by-default or lives only on the self-hosted GPU runner (test.nvidia_t4.yml, test.jetson_*.yml) does NOT protect a change merged through x86 CI.

2. Selectors, not just literals. Blocks/UQL dispatch on runtime values (PROPERTIES_EXTRACTORS in inference/core/workflows/core_steps/common/query_language/operations/detections/base.py). For every branch / enum value / property_name / kind added, a test must supply that runtime value through the block and assert its output. One test per new branch is the bar (PR #1129 added the anchor property_names plus one behavior test each).

3. Test isolation. Monkeypatching goes through the function-scoped monkeypatch fixture (auto-undone), never module-level assignment. Module stubs belong in a registered stub_plugins/ package, not sys.modules hacks. autouse fixtures must restore any shared global they touch.

4. Gate direction and GPU-unit routing. For any new skipif/env gate, grep the workflows for the var and confirm a job sets it so the test actually runs (e.g. SKIP_SAM2_TESTS=False in test.nvidia_t4.yml, SKIP_LMM_TEST in integration_tests_inference_server_x86.yml). A gate that defaults to skip and is never flipped is a dead test. GPU-only tests under inference_models/tests/unit_tests must carry gpu_only plus the dependency marker for the environment they require (for example, trt_extras). Confirm that .github/workflows/unit_tests_inference_experimental_gpu.yml installs that dependency group and selects the same marker intersection. The ordinary CPU unit workflow intentionally deselects every gpu_only test.

5. Marker correctness. Any new marker used in the diff must be declared in the relevant pytest.ini. inference_models/pytest.ini registers its hardware and dependency markers separately from the repository-root configuration. Unregistered markers silently no-op the intended deselection (PR #1724 had to add workflows).

6. Real coverage over mock theater. The assertion must observe the changed output, not a mock return value the test itself set. A test whose mocks reproduce the exact code under change asserts nothing about the change.

7. Assertion stability. Model-prediction tests assert shape/contract/tolerances, not brittle exact floats on numerically unstable ops (PRs #2026, #1646, #1394).

Example implementations (point authors here)

• tests/workflows/unit_tests/core_steps/formatters/test_property_extraction.py — canonical selector coverage: one test_property_extraction_block_with_<anchor> per runtime property_name, driven through the block, asserting real output (PR #1129).
• tests/workflows/integration_tests/execution/test_workflow_top_prediction.py — behavior-level execution test: full workflow compiled + run with $inputs.*/$steps.* selectors, covering non-empty and empty-detections paths.
• tests/workflows/integration_tests/execution/stub_plugins/ — isolated module stubs done right: fake blocks in registered plugin packages loaded per-test, not sys.modules monkeypatching.
• tests/inference/unit_tests/conftest.py (_patch_aioresponses_stream_writer, autouse=True + monkeypatch) — scoped, auto-undone patching; contrast with module-level global assignment.
• tests/inference/integration_tests/test_sam2.py (SKIP_SAM2_TESTS gate) + .github/workflows/test.nvidia_t4.yml (sets SKIP_SAM2_TESTS=False) — the gate-plus-runner pair for checking any new env-gated test has a CI job that flips it.
• inference_models/tests/unit_tests/models/rfdetr/test_triton_universal_preprocess_runtime.py (gpu_only + trt_extras) + .github/workflows/unit_tests_inference_experimental_gpu.yml — canonical GPU-only unit-test routing through a dependency marker.
• The relevant pytest.ini — repository-root and inference_models suites register different marker sets; add every new marker to the configuration governing its test path.

Reference PRs

• #1129 — selector coverage established (one behavior test per new property_name).
• #1724 — added the workflows marker to pytest.ini.
• #2526 — fixed a test left on an old run_workflow(...) signature after args moved to the WorkflowRunner constructor.
• #2168, #1539 — broken/missing test asset and stale stub after a dependency move.
• #2026, #1646, #1394, #1100, #1044, #1033 — loosened brittle exact-value prediction assertions.