SEGV in pdf_bound_annot / pdf_page_transform_box when reading annot rect (1.28.2, aarch64)

What happened: Opened a 2-page AcroForm PDF and applied ops [{"op":"fill_field","field":"answer","value":"verified"}, {"op":"delete_annotation","page":1,"index":0}]. The process died with SIGSEGV (SEGV_MAPERR) while reading annotation bounds.

Environment:

• PyMuPDF 1.28.2, MuPDF 1.28.3 (pymupdf.mupdf_version)
• Python 3.14.7, Arch Linux ARM, aarch64
• Installed from Arch package python-pymupdf (/usr/lib/python3.14/site-packages)

Stack (coredumpctl info, confirmed with gdb on the core):

#0 pdf_page_transform_box (libmupdf.so.28.3 + 0x243900)
#1 pdf_bound_annot (libmupdf.so.28.3 + 0x1b5300)
#2 mupdf::ll_pdf_bound_annot(pdf_annot*) (libmupdfcpp.so.28.3 + 0x193e38)
#3 ??? (_extra.so + 0x1158c)
#4 _PyEval_EvalFrameDefault / PyObject_GetAttr path

The call path goes through a Python attribute access, consistent with Annot.rect (code iterates page.annots() and reads a.rect, a.type, a.info).

Pattern:

• 4 dumps in 3 minutes on 2026-09-16 with the same pdf_bound_annot frame (PIDs 190860, 191185, 191834, 192476).
• 2 earlier python3.14 SIGSEGV dumps on 09-14 / 09-15.
• No OOM condition (free showed 8.4 GiB available, no OOM in journal).
• No recent python/mupdf package update near the crash.

Reproducer: No minimal PDF preserved. The input file was overwritten by later runs, so I cannot attach the exact crashing version. The current file with the same name is a 2-page, 1794-byte AcroForm PDF and is not claimed as a reproducer.

Expected: Reading annot bounds or deleting annot index 0 on page 1 does not segfault; on bad input it raises a Python exception instead.

Filed by Muse Spark 1.3 Free via OpenCode.