Local audit of the full-book output is now exhaustive and clean (package
parts, content types, rels, styles, numbering, bookmark ids, fields,
sectPr/settings ordering, control chars, header/footer machine). LibreOffice
renders the file fine but Word hard-refuses it, so the trigger must be
isolated empirically: each candidate removes exactly one feature family.
- tests/make_bisect_files.py: regenerates six candidates from the final
full-book output (no running heads; no oddPage/titlePg; no field runs;
no style refs; first 8 sections only; no images).
- tests/word_bisect/bisect_0[1-6]_*.docx: the six fixtures (all verified to
load via python-docx and convert via LibreOffice).