Instruction file imported from dertnius/sdd-semantic-pipeline (.github/instructions/tests.instructions.md). Copyright stays with the author.

Test conventions

These rules apply to tests/**.

• Default to the fast lane. A test with no marker must run with neither pandoc nor an ML model and stay fast. Run pytest -m "not slow" while iterating.
• Markers:
  • slow — needs the pandoc binary or a real embedding model. Mark any such test.
  • integration — full end-to-end with all services.
• Reuse tests/conftest.py fixtures rather than rebuilding state. Notably:
  • hashing_embedder — deterministic, model-free embedder. Inject it into SemanticPipeline(..., embedding_model=hashing_embedder) to exercise the real index→search path without a model (see tests/test_search_offline.py).
  • sample_document_model — a pre-built DocumentModel to chunk/embed.
  • the autouse PIPELINE_ENFORCE_WORKSPACE=false fixture bypasses the workspace guard; the contract itself is covered by test_workspace.py / test_cli_workspace.py.
• Prefer mocks for the vector store and embedder in unit tests; avoid network and large-model downloads in default paths.
• pythonpath=["src"] is set in pyproject.toml, so import sdd_pipeline works without installing.
• Test the contract, not internal incidentals — the offline search tests assert filters narrow / output is stable / hybrid runs, not semantic relevance (that lives in the slow e2e tests).
• A skill/asset wiring test should skip when the asset is absent (e.g. .claude/ is gitignored) — see test_docs_sync_skill.py.