M4.1c — Session reproducibility archive

[Ferrari-72]

Parent

Spawned from #8 (M4.1 tracking issue).

Objective

Archive each session's trajectory (the sequence of pred commands the AI ran) alongside the results file, so any leaderboard number can be re-validated offline.

Definition of done

• run_mini.py saves a trajectories/{safe_model}_{rule}.jsonl file containing each pred call (args + stdout + returncode) during a session
• build_index.py already picks up trajectory_file (done in M3.1 — Build the public leaderboard: dual metrics + certificate drilldown + trajectory #6) — no change needed there
• make test-unit still passes

Dependencies

Depends on #4, #6.

[Ferrari-72]

Completed — commit 4ddf544

What was done

run_mini.py

• save_trajectory(messages, path) — writes agent message history as JSONL (one {role, content} object per line); creates parent dirs automatically
• run_one(..., trajectory_dir: Path | None = None) — when trajectory_dir is set, saves to {trajectory_dir}/{safe_model}_{rule_name}.jsonl after the agent run; does nothing when None

build_index.py already picks up trajectory_file (added in #6) — no change needed there.

test_trajectory.py — 7 tests, @pytest.mark.judgment

Class	Test	Verifies
TestSaveTrajectory	test_creates_jsonl_file	file created
	test_one_line_per_message	line count = message count
	test_each_line_is_valid_json	each line parses as JSON
	test_each_line_has_role_and_content	each object has role + content
	test_creates_parent_dirs	deep nested path created automatically
TestRunOneTrajectory	test_trajectory_file_written	{safe_model}_{rule}.jsonl exists after run
	test_no_trajectory_when_dir_is_none	no file written when trajectory_dir=None

Test results

152 passed, 2 skipped   (make test-unit)