feat(test): coverage gate + close pass-1 gaps (test-strategy pass 1)

[silversurfer562]

Summary

Pass 1 of the test-strategy spec — Layer 3 (attune-rag, the contract source of truth for the attune product family).

Note: the single commit on this branch is titled `docs(specs): split dashboard-v0.2.0 + receive template-path-rename` due to how it was authored, but its diff is purely the test-strategy work below. PR title and description reflect the actual content.

• 87 new tests across 3 new files; total 437 passing + 3 xpassed
• Coverage: 75.88% → 79.94% (branch coverage; +4.06pp)
• Gate enforced at `fail_under = 77` (measure-then-lock at baseline-2pp)
• CI workflow gains coverage step on the ubuntu × py3.11 cell only (1 of 12 matrix cells); other cells run plain pytest for speed

What changed

Coverage gate

• `pytest-cov` configured in `pyproject.toml` with `branch = true`, `fail_under = 77`, and the standard `exclude_lines`.
• `tests.yml` runs coverage only on `ubuntu-latest × python-3.11` and uploads `coverage.xml` as an artifact.
• Stretch target for pass 2: 90% — rag is the API contract source of truth and warrants the highest gate of the four layers.

New test files

File	Purpose	Tests
`tests/unit/test_benchmark.py`	Closes the biggest gap (benchmark.py was at 10% line coverage). CLI exit codes, _load_queries paths, _percentile math, _run_benchmark aggregation.	19
`tests/unit/test_contracts.py`	Locks the public API: every `all` export importable + correct kind + has docstring. Pins `build_augmented_prompt` and `RagPipeline.run` signatures. Pins `CitedSource` consumer fields.	58
`tests/unit/test_providers_base.py`	LLMProvider Protocol runtime-check semantics, dataclass invariants, default `generate_with_citations` NotImplementedError.	10

Notes

• Task 7 (parametrize golden suite) was a no-op — the existing `tests/golden/test_golden.py` is already correctly table-driven via `pytest.param`. The audit's "hardcodes 40 queries" finding was stale.
• New `tests/README.md` documents how-to-run, the LLM-mocking pattern (attune-author conftest is the reference), the public-API contract policy, and the golden suite policy.

Test plan

• `pytest --cov` locally — 437 passing + 3 xpassed; 79.94% branch coverage
• `ruff check` clean on new test files
• CI matrix passes on ubuntu/macos/windows × py3.10–3.13
• Coverage artifact uploads correctly on the py3.11 cell

Out of scope (deferred to subsequent passes)

• `dashboard/show.py` (172 LOC, 0% covered) — biggest remaining gap; pass 2.
• `dashboard/refresh.py` and `eval/bench_prompts.py` partial coverage — pass 2.
• LLM/API-call discipline across layers — pass 2.
• Slow / flaky test handling, CI duplication audit — pass 3.

🤖 Generated with Claude Code