The mini opens with 3 byte-identical fields while the full shell fails even
with 3 (bisect_12) or 1 target (bisect_11), so the co-factor is outside the
TOC. Six new candidates keep fields and strip the remaining families or
probe the field mechanism:
- 13 drop scene-break paragraphs
- 14 drop all bookmarkStart/End (fields keep, targets vanish)
- 15 full shell with the mini's own 3-entry TOC spliced in verbatim
- 16 drop chapter title paragraphs (WAChapter)
- 17 rewrite PAGEREF instructions to PAGE (structure unchanged)
- 18 remove w:dirty (no auto field update on open)
bisect_03 (fields removed) was the only round-1 file Word opened, and
document.xml becomes byte-identical between the full book and bisect_03
once field runs are removed — but the mini book opens with identical
PAGEREF fields, so the fields interact with a second factor. Add six
candidates that keep the 28 PAGEREF fields while peeling other layers:
- 07 drop deep-copied chapter prose (keep titles/bookmarks/fields)
- 08 drop numbering properties
- 09 inline hyperlink wrappers
- 10 remove the whole TOC section (cleaner no-fields control)
- 11 rewrite all PAGEREF targets to a single bookmark
- 12 keep fields only for the first 3 TOC entries
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).
Root cause found by comparing the full-book output (Word refused) with the
new mini-book fixture (Word opened): the output's styles.xml was replaced
by the template's, yet deep-copied manuscript paragraphs still referenced
four KDP paragraph styles (WorstChapter, WrstNrmlChapText, WrstSection,
WAChpNrm) and the Hyperlink character style. Word validates every
referenced styleId and refuses files with dangling references; LibreOffice
silently falls back to Normal.
- Add _merge_missing_styles(): copy any w:style referenced by the output
body but missing from styles.xml out of the source manuscript, resolving
basedOn/next/link chains. Runs at the end of Pass 2.
- Fix TOC entry construction to use paragraph_format.tab_stops (python-docx
inserts w:tabs in the schema-correct CT_PPr position instead of appending
after w:jc/w:spacing).
- Regression tests: test_style_references_resolve (no undefined pStyle/
rStyle) and test_body_pPr_children_schema_ordered (CT_PPr sequence).
55 tests pass. Full-book output regenerated (332 pages; merged styles
present; 0 undefined style refs; 0 pPr violations).
tests/make_word_test_files.py regenerates three fixtures:
- tests/word_test_input.docx - tagged mini manuscript (book start through
req review, all section roles, <<<>>> scene breaks)
- tests/word_test_output.docx - output of format_document() on the above
- tests/python_docx_control.docx - untouched python-docx save (baseline)
Purpose: bisect the Word-open failure. The full manuscript output is
structurally clean by every OOXML check (package rels, content types,
sectPr/settings ordering, header/footer parts, fields, hyperlinks, custom
parts all validate); the fixtures let Word itself tell us whether the
pipeline or the manuscript content is the trigger.
sectPr children were emitted with headerReference/footerReference after
pgSz..mirrorMargins (SECTPR_ORDER), and w:settings had evenAndOddHeaders /
autoHyphenation appended after decimalSymbol/listSeparator. LibreOffice
tolerates out-of-order XML; Word validates CT_SectPr and CT_Settings
strictly and refuses the file.
- Move headerReference/footerReference to the head of SECTPR_ORDER per
ECMA-376 CT_SectPr.
- Add SETTINGS_ORDER + _reorder_settings() using the full CT_Settings
sequence (mathPr in the m: namespace) and apply it in Pass 2.
- Add regression tests asserting monotonic CT_SectPr and CT_Settings
order for every section. 53 tests pass.
- Spec §4.1: running headers carry titles only (recto: chapter/section
title; verso: author name + book title); page-number fields removed
from headers.
- Spec §4.2: page numbers centered in the footer on both verso and recto
pages; front-matter display pages (copyright, dedication, epigraph,
TOC) show their Roman numeral on the opening page too.
- Spec §2.2: lowercase Roman numbering visible from the copyright page
through the front matter; Chapter 1 restarts at Arabic 1.
- Chapter/back-matter sections keep running matter on body pages; only
the first page of each chapter is suppressed (spec §4.3).
- Drop titlePg on single-page front-matter display sections so footer
page numbers render in LibreOffice and Word alike.
- Preserve <<<>>> scene-break cues (centered, 12pt before/after) and
first-word-only capitalization of chapter openers.
- UI: remove the obsolete "Page Numbers" selector; footer_mode is
accepted-but-ignored for backward compatibility.
- Tests updated to the new spec behavior (51 pass).
