Context

Part of the Testing Infrastructure & Strategy Overhaul epic (issue 1a). This is the foundational taxonomy that other issues in the epic depend on.

See Discussion #711 for full background and the marker inventory.

Objective

Define and implement granularity markers for the test suite, classifying tests along the agreed taxonomy:

• unit — tests individual functions/classes in isolation
• integration — tests full stack using mocked backends
• e2e — tests against real backends, no mocks
• qualitative — subset of e2e that asserts on LLM output quality (non-deterministic)

Work

• Register the new markers in conftest.py and pyproject.toml pytest config
• Decide on the convention for unit tests: either "no marker = unit" (implicit, low friction) or an explicit @pytest.mark.unit (self-documenting, enables pytest -m unit). If explicit, add validation that rejects unmarked tests (e.g. conftest check or CI lint step)
• Apply granularity markers across all test files (80 files total; 30 currently unmarked)
• Implement tiered timeouts per granularity (~30s unit, ~120s integration, ~600s GPU/e2e) to replace the current single 900s global default
• Update contributor documentation with the new marker conventions

Notes

• The two-dimensional taxonomy (granularity + backend) was agreed in the March 23rd planning call
• Backend and resource markers are audited separately in issue 1b
• qualitative is a subset of e2e, not a separate tier