feat(html): honour the manual page break, and stop calling the automatic one a break

[andiwand]

🤖 Generated with Claude Code

Steps 1 and 2 of the plan in #174 (comment).

The signal was wrong

ElementType::page_break came from ODF's text:soft-page-break, which [OpenDocument] 5.1.1 makes a record of where the producer's layout engine broke pages — informational, not authored. In about.odt one sits mid-heading, between a draw:frame and the word "Features". Rendering it would re-impose LibreOffice's pagination for LibreOffice's page size on a viewport that has nothing to do with it. It was 240 of the 312 elements #805 measured the renderer dropping. It is no longer parsed.

The real break was not modelled

format	manual break	before	now
ODF	fo:break-before / fo:break-after on the paragraph style	✗	✓
OOXML	w:pageBreakBefore in w:pPr	✗ (docx produced no break at all)	✓
RTF	\page	✓ element	✓ element
.doc	\x0C	✓ element	✓ element

New BreakType { none, page, column } and ParagraphStyle::break_before/break_after. The explicit none is load-bearing: sample1.docx's TOCHeading is w:basedOn="Heading1", which breaks, and says <w:pageBreakBefore w:val="0"/> to stop — so "off" has to be told from "unsaid" or the derived style inherits a break it turned off.

What the renderer does

The text view wrote one .odr-page-outer sheet. It now closes it and opens the next at a manual break, from either source. This is not pagination: nothing computes where a page ends, so the boxes are of unequal height. A box means "what the author put on a page of its own", never "what fits on a page" — which is why it does not wait for #76, it sidesteps it. A document with no manual break is the one box it always was, and .odr-pages already stacks and centres them with a 16px gap, so no CSS changed.

break-before:page goes out on the paragraph alongside, which is what print and PDF export need. BreakType::none writes nothing — it is the CSS initial value.

mixed-layout.odt, the fixture the JNI, Apple and wasm suites share, now renders as the four sheets it is:

sheet 1: Portrait 1    sheet 2: Portrait 2    sheet 3: Landscape 1    sheet 4: Portrait 3

Verification

• core: full suite, 1399 passed / 6 skipped, and the emitted tree matches the regenerated reference exactly. Nine new tests: the ODF ones build flat-ODT inline (style read, auto → none, the split, no empty leading sheet, soft break is not content, reflow gets css only); the OOXML one proves the on/off inheritance chain.
• JNI / Python / Apple: all suites pass; BreakType and the two fields are mirrored in each.
• feat(html): warn on the element types the renderer drops, instead of dropping them silently #805's dropped-element warning now reports zero across the corpus, down from 312.

Reference output

39 files changed, all odt/docx/doc — nothing else moved. Pins advanced to 3dcc4fb / 9e9cc58.

• Their text is byte-identical after tag stripping, all 39.
• 13 gained page boxes (785KB.doc 1→65, Vektoranalysis Zusammenfassung.odt 1→107).
• 14 gained a css break where the box could not be split. Checked through the Python bindings that in every one the break is either on the very first block (where splitting would open an empty sheet) or nested below the text root: zero cases where a top-level break failed to split.
• 12 differ only in the data-odr-path sibling indices the removed soft breaks used to occupy. No pixel or behaviour change — frontend.cpp selects on the attribute's presence, never its value.

Not in this PR

• Step 3: docx's inline <w:br w:type="page"/> (66 corpus occurrences) sits inside a run, so honouring it visually means splitting the paragraph. break-before:page from step 1 already handles it for print.
• A break may also switch master page, so mixed-layout.odt genuinely has two page geometries where text_root_first_master_page gives us one. Per-sheet page layout is its own issue.