v3.38.0

3.38.0 (2026-09-12)

Phase 19 B: the audit-depth batch (one class per commit, each with its fixture and false-positive note)

• Content-duplicate fingerprinting (scripts/audit_duplicates_content.py): 64-bit simhash over 3-word shingles of each book's spine text finds re-downloads filed under different metadata; a bottom-32 shingle-sketch candidate pass catches omnibus containment (an omnibus's simhash is not close to the standalone's, so Hamming distance alone never sees it), and candidates are classified exactly: near_duplicate vs omnibus_overlap. Front-matter-only files are excluded (under 200 shingles a simhash is noise) and formats without an extractor are skipped, not guessed; legitimate public-domain reissues are findings by charter. ~1s per book: scope with --search/--ids.
• Truncation cross-checks (scripts/audit_truncation.py): the Count Pages plugin's page counts meet poppler's pdfinfo; more than 20% disagreement is page_count_mismatch, and stale_plugin_data (catalogued size drifted, the plugin's own needs_scan flag, post-scan mtime) is its own class. Only PDF rows are checked: the plugin's EPUB pages are word-count estimates by design.
• Author-sort sanity in validate_metadata.py: AUTHOR_SORT_NOT_INVERTED (sort identical to a multi-word display name) and AUTHOR_SORT_ORPHAN (matching no legitimate shape of the book's authors). Hosted locally by decision: cquarry never boxed the predicate, and the promotion remains a future option. The first cut compared against display names only and flagged 7718 of 7842 real books; the real-library probe caught it within minutes and the fix accepts every legitimate Calibre shape (display name, the authors.sort column, mechanical inversion, the &-joined multi-author sort). Real probe after the fix: 0 orphans.
• DB-level ISBN checksum in validate_metadata.py: INVALID_ISBN warns on any isbn identifier failing its check digit (cquarry's existing helper; no file open). Blank-ish values are "no ISBN", not invalid ones.
• Cover aspect-ratio bands (scripts/audit_cover_aspect.py): covers outside w/h 0.55-0.80 report as advisory cover_aspect_narrow/cover_aspect_wide through cquarry's header-only image readers (no image library). Legitimate landscape art exists; the audit surfaces the distribution, the operator judges.
• FTS coverage audit in --audit: the A.1 staleness classes render as fts_coverage CSV rows. With no sidecar the CSV stays silent (the rows would be the whole library) and the prose summary carries the absent-sidecar note instead.
• PDF battery depth in check_pdf.py: text sampled at pages 1, middle, and last (the page-1-only sample passed OCR-once scans clean; a partial pass is now text_layer_partial), and image DPI parsed from the existing pdfimages -list output as an area-weighted mean (a small sharp logo cannot hide a full-page 72-dpi scan; below 150 dpi is a low_dpi advisory). Advisory classes only: the structural total and the run-verb seam contract are unchanged.
• Copyright year vs pubdate in audit_isbns.py: the front-matter pass captures (c)-years (each copyright marker claims the years on its line), and an earliest-year-vs-pubdate gap over two years is a year_mismatch advisory with its own report section, riding the exit-1 findings contract. A reprint legitimately prints the original year; the report asks whether the pubdate describes this edition.
• Skills sync: swept both import skills for the new audit surface; nothing teaches the old shapes, no staleness found. Suite: 433 → 461 tests.