Cross-page paragraph/word stitching (deterministic, Path B)

[Arkoniak]

Join paragraphs and words split across page boundaries — now at HTML level, inside build_html.py.

Depends on: #30 (enriched voted_tokens per page; headers already stripped from body_text
and label = page_header/page_footer is already voted — no pre-processing needed here).

The original plain-text stitch.py approach is superseded: operating on the structured token stream
(with paragraph_start and label attributes) is cleaner and avoids the header-wedging problem
described in brainstorm Layer 4.

Logic (unchanged):

• Word split: last token ends with ASCII - + first token of next page is lowercase → join without hyphen
• Paragraph continuation: no terminal punctuation (.!?") + first token lowercase → join with space
• Otherwise: new paragraph break (</p><p> boundary)

Distinguish — (em-dash, not a word-break) from - (ASCII hyphen at line end).
Unlimited multi-page mode as optional cross-check signal for flagging uncertain boundaries.

See brainstorm Layer 4.

[Arkoniak]

Implemented as part of #30 inside scripts/build_html.py.

All three cases from the issue spec are covered:

• classify_boundary() — word split (ASCII hyphen + lowercase first word on next page), paragraph continuation (no terminal punctuation + lowercase), new paragraph otherwise; correctly distinguishes - from —
• stitch_boundary() — merges element lists at the page boundary, strips the trailing hyphen for word-split joins
• build_book_html() — applies cross-page stitching in the page iteration loop, with gap detection for non-consecutive pages

Stitching decisions are logged to stitching_decisions.json alongside book.html.