test: add integration tests with in-memory Phoenix - #36

Tests

Uploader (5 tests)

• Creates one dataset per unique tag
• Datasets have correct example counts
• Skip mode is idempotent (no duplicates)
• Append mode adds examples to existing datasets
• Custom input/output keys are preserved in examples

Runner (7 tests)

• Dry run completes without error
• Full run returns experiment results
• Evaluator execution during run
• Name filter selects specific experiments
• Missing dataset exits with code 1
• Concurrent execution (concurrency=2)
• Custom prefix and metadata

Architecture

• Session-scoped Phoenix fixture: one server per test session on a random free port
• Unique dataset names (uuid suffix) to avoid conflicts between tests sharing the same server
• @pytest.mark.integration marker: tests are excluded by default (addopts = "-m 'not integration'" in pyproject.toml)
• CI job: dedicated test-integration job added to ci.yml, included in the ci-success gate
• Makefile: make test-integration target

Running

make test-integration         # run integration tests only
make test                     # unit tests only (default, skips integration)
uv run pytest tests/ -m integration   # manual