Pages: pre-warm matplotlib font cache in parent before spawning subprocesses

The font cache build takes 30-60 s on a fresh CI runner inside
`import matplotlib.pyplot as plt`.  When this long initialization coincides
with Metal GPU loading (_m1_gpu_ops) in the same process it causes SIGSEGV.

Fix: run the matplotlib import and a dummy figure in the parent (orchestrator)
process which never touches Metal.  The cache is written to disk so every
notebook subprocess finds it already built, making plt's import instant and
eliminating the Metal ↔ matplotlib initialization overlap.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: larsgeb

scripts/render_notebooks.py

@@ -160,6 +160,22 @@ def main() -> None:
     notebooks = [Path(p) for p in sys.argv[2:]]
     out_dir.mkdir(parents=True, exist_ok=True)
 
+    # ------------------------------------------------------------------
+    # Pre-warm the matplotlib font cache.
+    #   Building the cache takes 30-60 s on a fresh CI runner and happens
+    #   inside `import matplotlib.pyplot as plt`.  When this coincides with
+    #   Metal initialisation in the same subprocess the process can crash
+    #   (SIGSEGV).  Running the import here — in the parent, which never
+    #   touches Metal — writes the cache to disk so that every notebook
+    #   subprocess finds it ready and skips the slow build entirely.
+    # ------------------------------------------------------------------
+    print("Pre-warming matplotlib font cache ...", flush=True)
+    import matplotlib.pyplot as _plt  # noqa: PLC0415
+    _plt.figure()
+    _plt.close("all")
+    del _plt
+    print("  font cache ready.", flush=True)
+
     # ------------------------------------------------------------------
     # Phase 1 — execute each notebook in an isolated subprocess.
     #   If the subprocess crashes (e.g. SIGSEGV / exit 139), we write a