fix: run notebook smoke from the workspace root so dataset paths resolve

.github/scripts/run_smoke.py

@@ -101,6 +101,13 @@ def execute_notebook(nb_path: Path, env: dict) -> tuple[int, str]:
     # Write the executed copy to a throwaway path so the on-disk notebook
     # under notebooks/ is never modified — checked-in notebooks stay clean.
     tmp_dir = Path(tempfile.mkdtemp(prefix="smoke_nb_"))
+    # `jupyter nbconvert --execute` runs the kernel in the *notebook's own
+    # directory*, but the workspace convention (and the script smoke, which runs
+    # with cwd=WORKSPACE) is that relative `dataset/` paths resolve from the repo
+    # root. Stage a temporary copy of the notebook at the workspace root so the
+    # kernel's working directory is the root and root-relative paths resolve.
+    staged = WORKSPACE / f".smoke_run_{os.getpid()}_{nb_path.name}"
+    shutil.copyfile(nb_path, staged)
     try:
         result = subprocess.run(
             [
@@ -113,14 +120,15 @@ def execute_notebook(nb_path: Path, env: dict) -> tuple[int, str]:
                 str(tmp_dir),
                 "--output",
                 nb_path.name,
-                str(nb_path),
+                str(staged),
             ],
             cwd=str(WORKSPACE),
             env=env,
             capture_output=True,
             text=True,
         )
     finally:
+        staged.unlink(missing_ok=True)
         shutil.rmtree(tmp_dir, ignore_errors=True)
     return result.returncode, result.stdout + result.stderr