fix: regenerate_notebook resolves the source by relative path, not filename

[Jammy2211]

Summary

#261 looked the source script up as scripts_dir / nb_path.name, which drops
the subdirectory. Every workspace notebook lives in one, so
notebooks/imaging/model_fit.ipynb searched for scripts/model_fit.py instead
of scripts/imaging/model_fit.py.

The stale-notebook recovery was therefore dead on arrival for real
workspaces: FileNotFoundError, caught, the original FAIL left standing. No
false PASS — but a feature that silently did nothing.

Worse than not finding it, a bare filename can find the wrong one. Two
model_fit.ipynb under different topic folders map to two different scripts,
and a same-named script at the scripts root would be picked up for all of them.

The fix

execute_notebook gains notebook_rel; execute_notebooks_in_folder computes
it with file.relative_to(Path.cwd() / directory), so the folder runner always
passes it. A direct caller that omits it keeps the old bare-filename fallback,
which is correct for a notebook sitting at the root.

Why the existing tests missed it

They used a flat fixture — notebook and script both at the root — which is
exactly the one layout where this bug is invisible. That is the real lesson
here: the fixture was shaped like the happy path rather than like production.

The two new tests use a nested notebook plus a decoy script at the scripts
root, so resolving by filename picks the decoy and fails loudly rather than
merely not finding anything.

Test Plan

• pytest tests/ -q → 391 passed, 5 skipped, 0 failed
• Both new tests negative-tested: they fail against the old resolution
  and pass against the fix (this repo's stated convention for guard tests)
• Tenant-firewall gate → OK

Found while preparing the phase-2 workspace delegators (PyAutoHands#260), before
any of the three notebook-capable workspaces was touched.

Generated by the PyAutoLabs agent workflow.

---

Generated by Claude Code