feat(db): line_dh — dibbur-hamatchil index

[palmoni5]

Resolves the DB side of the forum request: searching Rashi/Tosafot by dibbur hamatchil, and rendering dibburim as sub-headings.

Problem

A commentary's dibbur hamatchil is embedded in each line's content, and the TOC stops at daf/perek level (Rashi on Berakhot: 126 TOC entries for 3,265 lines). There is no way to search a commentary by its opening words, or to navigate a daf's comments by dibbur.

What this adds

line_dh(bookId, dhText, lineIndex) — normalised text, not a hash (unlike line_ref), so the client can prefix-match while the user types, straight off the WITHOUT ROWID PK btree (dhText >= ?prefix AND dhText < ?prefixEnd), and render the rows as virtual sub-headings — without touching tocEntry/tocText.

Two shapes cover the corpus, one per book:

• <b>…</b> prefix — Rashi on Tanakh, Mishnah commentaries, most Otzaria-sourced books.
• Text before the first spaced dash — Sefaria's Talmud commentaries (עד סוף האשמורה הראשונה – שליש הלילה…), including the daf's first comment that ends with . instead of a dash.

False-positive defences (measured on a real DB)

Bold ≠ dibbur in general: 2.0M lines open with <b>, including discourse-word emphasis (והנה ×28K, אמנם ×10K), structural markers (מתני', גמרא, בא"ד), and whole-line-bold decorated headings (57K). Three layers keep them out:

1. Book-dominance gate (the main one): a book is indexed only when one shape covers ≥40% of its content lines with ≥10 hits. Books that merely bold an occasional word or use an incidental dash never reach the threshold.
2. Line guards: real commentary text must follow the dibbur (whole-line bold is a heading), length ≤100 chars, <h1>–<h6> excluded.
3. Structural-marker stop-list, compared before quote-stripping — so the locator תוד"ה is dropped while the genuine dibbur תודה survives.

Parity

The stored value is DhKey.normalize() output; the Otzaria client must normalise the typed query with the same rules. Both sides pin to dh_key_fixtures.json (DhKeyParityTest here), same pattern as ref_key_fixtures.json.

Verification

• Regression coverage for extraction, normalisation, base-text exclusion, atomic rollback, and manifest compatibility — passes.
• Full build over a real 7GB seforim.db: 1,919,349 dibburim over 3,435 non-base/dependent books in 49s; pure base texts contribute zero rows.
• buildLineDhIndex runs after every book-writing stage and is idempotent (rebuilds from scratch).

🤖 Generated with Claude Code

Review hardening

• Rebased directly onto otzaria; duplicated line_ref commits were removed.
• Pure base texts are excluded, while dependent works marked as base remain eligible through book_base_text.
• Structural markers handle Hebrew, ASCII, doubled-apostrophe, and curly quote variants.
• Rebuilds are transactional and roll back to the previous index on failure.
• Schema-4 manifests now emit patchFormatVersion, matching the updater contract.