CI runs no tests; 20% of package never imported by the test suite

[thedataengineer]

CI runs zero tests. .github/workflows/publish.yml is the only workflow and it gates version sync, python -m build + twine check, and a wheel import smoke test. No pytest job exists, there is no make test target, and pytest is not a declared dev dependency even though tests/test_worker_hub.py and tests/test_remote_pool_logs.py import it.

Coverage state of the CPU suite:

• 3,473 lines (about 20% of the package) are never imported by any test: backends/torch.py (1,141 lines, the production training path), backends/mlx.py (1,125), cli.py (564), backends/verl.py, accel.py, bottleneck.py, charts.py, ascii.py, _cli_entry.py, _quiet.py.
• The public SDK surface has zero assertions: models.py (load/finetune/generate/chat/save, _parse_reply), training.py resolvers (resolved_target_modules, resolved_warmup, resolved_eval_steps, resolve_total_steps) and TrainingRun.to_dict/from_dict (the serialization boundary the hub round-trips jobs through via serve.py:137), backends/__init__.py:37 select_backend, methods/base.py:50 validate_base (the quantization legality guard), capture.py:200 trajectories() (episode reconstruction), and data.py format detection plus as_chat/to_texts (the render path backends train on; silent misdetection corrupts training data with no error).
• tests/gpu/test_cuda.py exits 0 when every test skips (main() at line 338 returns 1 if fails else 0), so on a non-CUDA box the GPU suite "passes" having run nothing.

Cheapest fixes first: the training.py resolvers, select_backend, cli.py _coerce/_apply_set/_load_config_file, and data.py _detect_format are pure functions with no heavy deps. A pytest job in CI plus tests for those closes most of the exposure in one pass.

[thedataengineer]

Addressed in ac6e332: added .github/workflows/test.yml (pytest on 3.10 + 3.12, every push and PR, plus compileall), a make test target, and a [dev] extra declaring pytest. New tests cover the TrainConfig resolvers + dict round-trip, select_backend's five branches and four install messages, and data.py format detection + the as_chat/to_texts/split render path — 181 passing locally.

Two fixes were needed to make the suite runnable without a training backend installed: run_worker probed select_backend for the registration name even when handed a backend_factory, and test_more_plus_merge imported torch at module scope. Also, tests/gpu/test_cuda.py now exits 2 when every test skipped instead of 0 (exiting 0 read as 'GPU suite passed' when nothing ran).

Still open: backends/torch.py and backends/mlx.py (2,266 lines) remain untested outside the GPU box, and models.py/cli.py/capture.py have no assertions yet.