Page-level classifier: content / chapter_opener / special_page / picture

[Arkoniak]

Background

build_html.py currently treats all non-picture pages as stitchable content and renders them as <p> paragraphs. This is wrong for:

• Epigraph / front-matter pages (short isolated quote, e.g. page 8 "Law stops at atmosphere") — should be <section epub:type="epigraph"> or <blockquote>, NOT stitched to neighbours
• Chapter openers — should trigger a new EPUB chapter, NOT stitch with the preceding page (this mostly works via the heading label already, but needs to be made explicit)
• Picture pages — currently skipped entirely; should emit <figure><img> (see Image extraction from PDF: extract_images.py for picture-page embedding #34)

Brainstorm: docs/brainstorm.md § "Page-level classification for HTML assembly (design 2026-06-29)".

Page classes

Class	Detection	HTML output	Stitching
content	default	<p>	yes
chapter_opener	first token block label = title/heading	<h1>/<h2> + <p>	no (starts new chapter)
special_page	total_tokens < 40 AND no title/heading AND not picture	<blockquote> / <section epub:type="epigraph">	no
picture	picture_page=True	<figure><img src="images/page_NNN.jpg"/></figure>	no

Plan

1. Add classify_page(page_result) -> str function in build_html.py — pure heuristic, no model needed, operates on voted_tokens + picture_page flag already in consensus output.
2. Update build_book_html() to call it per page and route to the appropriate renderer.
3. Picture pages: emit <figure><img> (requires images to be extracted, see Image extraction from PDF: extract_images.py for picture-page embedding #34).
4. Special pages: emit <blockquote> with no stitching on either boundary.
5. Chapter openers: block stitching with preceding page (already works via stitch_boundary label check, but make it explicit via page class).

Depends on

• Enrich consensus with voted_tokens + build per-page HTML fragments #30 (voted_tokens in consensus output — already done)
• Image extraction from PDF: extract_images.py for picture-page embedding #34 (image extraction — for picture page rendering)

Acceptance criteria

• classify_page() function with tests covering all four classes
• Epigraph page 8 ("Law stops at atmosphere") renders as <blockquote> in book.html
• Picture pages emit <figure><img> pointing to extracted images
• No regression on stitching stats (word_split / paragraph_continuation counts unchanged)