[kvnloo]

Issue

Refs #255

Changed files

• tests/test_julia_worker.py (new): isolated pytest fake packages and subprocess JSONL harness; 11 protocol tests.
• No production, fixture-directory, configuration, dependency, lockfile, or CI changes.

Acceptance checklist — tests implemented, not executed

• test_status_before_load — status before load; configured CLI flags, defaults, PID, and unloaded state.
• test_predict_round_trip — request ID, state/questions passthrough, default loader kwargs, latency/load timings, device.
• test_warm_then_predict_loads_once — warm/status/predict sequence and one resident model load.
• test_invalid_json_does_not_kill_worker — malformed JSON error, ignored blank line, continued status response.
• test_shutdown_stops_processing — early clean exit without extra stdout.
• test_request_error_is_reported_per_line — missing questions KeyError, bounded traceback, successful following request.
• test_missing_answers_become_empty_dict — empty model response normalizes to empty answers.
• test_stdout_carries_only_protocol_lines — parseable JSON on stdout; fake-library chatter on stderr.
• test_autocast_off_replaces_torch_autocast — off vs auto class names across independent processes.
• test_warm_on_start_failure_is_recorded — model-load exception saved in status, subsequent predict error, worker survives.
• test_missing_model_dir_exits_2 — missing required argument and invalid --autocast value exit 2 with usage errors.
• Tests import only Python stdlib + pytest and never import the worker in-process; fake julia and torch live under tmp_path, injected via subprocess PYTHONPATH.
• No real Julia, torch, HOME, network, sleeps, or additional dependencies in subprocess environment; 30-second subprocess timeouts.
• Only the issue-allowed new test file changed.
• Maintainer/CI to run tests and verify outcomes.

tests NOT RUN - ChatGPT cannot execute code (written via GitHub connector; no pass counts or CI claims).

Exact test commands

pip install -e '.[test]' && python -m pytest -q tests/test_julia_worker.py

Notes

The worker source on dev already matches the specified status, warm, shutdown, predict, and load-error contracts. No production bug fix was justified.