fix: run notebooks with the kernel cwd pinned to the workspace root

[Jammy2211]

Notebooks execute with the kernel's working directory pinned to the workspace root, so root-relative auto-simulate guards resolve.

Closes #204.

The bug

jupyter nbconvert --execute starts the kernel in the notebook's own directory, not the directory nbconvert was launched from. The workspaces document the opposite (autolens_workspace/AGENTS.md: "Scripts are run from the repository root so relative paths to dataset/ and output/ resolve correctly"), and the auto-simulate guards depend on it:

if al.util.dataset.should_simulate(str(dataset_path)):
    subprocess.run([sys.executable, "scripts/imaging/simulator.py"], check=True)

That path resolves from a script run at the root and cannot resolve from a notebook run in notebooks/<topic>/ — exit 2, "can't open file". It is the single largest contributor to the red run_notebooks shards. The .py scripts were always fine, which is why run_scripts mostly passed.

Why this shape

Both options in the original issue turned out to be unimplementable as written, so I checked each before building:

• "Force the kernel's working directory" is not a cwd= kwarg. The launcher's cwd is already the root; nbconvert overrides the kernel's cwd regardless and exposes no CLI flag for it. The only knob is resources['metadata']['path'], which nbclient turns into the kernel cwd (nbclient/client.py:535) — Python API only.
• "Resolve off __file__" cannot work: __file__ is not defined in a notebook kernel (verified). It exists in the .py scripts, which already work, and is missing in exactly the broken context.

So execute_notebook now subprocesses a new autohands/run_notebook.py that sets the resource path. The subprocess boundary is kept deliberately — process isolation, BUILD_SCRIPT_TIMEOUT and the per-notebook environment are all unchanged. Only the kernel's cwd moves. No tutorial scripts are touched.

The load-bearing constraint

build_util.is_clean_skip_exit parses the run's combined stdout/stderr for a CellExecutionError marker whose last line is SystemExit: 0, to tell an intentional optional-dependency skip from a genuine failure. run_notebook.py therefore leaves CellExecutionError uncaught, so Python's own traceback reproduces that exact shape. A handler that reformatted the message would silently break skip detection — noted in the module docstring.

Verified against the classifier:

[OK] skip.ipynb  exit=1  is_clean_skip_exit=True  (want True)
[OK] fail.ipynb  exit=1  is_clean_skip_exit=False (want False)
[OK] err.ipynb   exit=1  is_clean_skip_exit=False (want False)

Verification

End-to-end on a real workspace notebook (autogalaxy_workspace/notebooks/guides/plot/examples/mat_plot.ipynb), whose guard shells out to a root-relative simulator:

OLD PATH: jupyter nbconvert --execute   -> exit=1  CalledProcessError
NEW PATH: run_notebook.py               -> exit=0  PASS

Two regression tests added (tests/test_run_notebook_cwd.py) covering the kernel cwd and the root-relative-subprocess case. Full suite green: 236 passed.

🤖 Generated with Claude Code