content audit sec01-06: fixes (audit_content.py + 12 file/card corrections)

- tools/audit_content.py: per-file vs item-title triage (SUSPECT/SMALL/NO-TEXT/NO-CYR)
- sec01: .part removed; 10 files restored .pdf ext (LFS rename); #04/#19/#27 verify notes
- sec02 #16: corrupted double-encoded txt -> clean flib fb2 b/844509; #17/#22 doc verified via catdoc
- sec03 #11: -b epub = Spanish von Franz (Paidós 1983, forged EN OPF) -> bonus label
- sec03 #42: КАРО 2012 'Irish Tales' = EN reader (not RU, not the listed book) -> bonus rename
- sec03 #56: Onians b/c/d/e = Cambridge 24-87KB previews -> labeled
- sec06 #18: filename 1981->1989 (Perera 'Descent to the Goddess' 1989; 1981 = other book)
- sec06 #28: ia Zimmer OCR layer = foreign Devanagari text; images = RKP (p.100 verified)
- sec04 #30: buksmart 2020 = RU/FR bilingual w/ VeryPDF watermarks -> note
- linter: azw3/mobi ext + underscore in fname regex; manifest cross-check ignores off-disk rows;
  8 legacy cards section-order fixed (CR before DL)
- AGENTS.md: 'Content audit (2026-09-25)' lessons section
- check-md: 0/219

AGENTS.md

@@ -531,6 +531,30 @@ sections/0N-<name>/  # one .md per BOOK (chapter-level list items are merged int
    queue bash (dual-queue downloads corrupt the cookie jar). AA (annas-archive.gl) = 403 bot-wall
    for anon downloads from this box; i2p not worth the router setup.
 
+## Content audit (2026-09-25) — what the full-file vs item-title sweep found
+
+Tool: `tools/audit_content.py <sec>...` (samples first/mid pages of every file, scores title
+overlap, flags SUSPECT/SMALL/NO-TEXT/NO-CYR). Ran on all sections 01–06 (sec05: fix the
+per-item header parse first — sec05 uses «— ✅» not «— N files»). Findings (commits
+`85c41ac` + follow-ups):
+1. **Mislinked whole files caught:** sec01 #29 RU fb2 = fantasy novel; sec01 #47 two review
+   clippings (1950 «About Books»); sec03 #11 `-b` epub = Spanish von Franz (Paidós 1983) with
+   forged EN OPF title; sec03 #42 «КАРО 2012» = English reader (Irish Tales), not the listed
+   book and not a RU translation; sec02 #16 bosnak .txt = double-encoded mojibake (replaced by
+   clean flib fb2 b/844509).
+2. **Foreign OCR layer (file OK, text layer wrong):** sec06 #28 ia Zimmer = all-Devanagari OCR
+   (Lal Bahadur Shastri Academy stamp); images verified = RKP «Philosophies of India» (p.100
+   «PART II Philosophies of Time»). Keep; the twin lg file is the clean copy.
+3. **Previews/fragments labeled as editions:** sec03 #56 Onians b/c/d/e = Cambridge 24–87 KB
+   chapter previews (marked ПРЕВЬЮ in MANIFEST); sec01 #12 was a 23-pp PagePlace preview (→ full).
+4. **False positives to expect:** windows-1251/UTF-16/KOI8 fb2/txt (decoding, not corruption),
+   image scans without OCR (NO-TEXT/NO-CYR — verify by rendering pages: pdftoppm/ddjvu +
+   imagemagick), spaced-out OCR titles («T H  C  T  F»), library stamps (Cornell/India/Claremont),
+   VeryPDF watermark pages (sec04 #30 buksmart — watermark + FR part, legit 2020 ed.).
+5. **Bonus rule in practice:** different-but-related books stay as bonus files with
+   `Бонус: ДРУГАЯ книга (...)` note, status unchanged; rename file to drop the false lang/year
+   label. Same for wrong-year filename (sec06 #18 «1981» → 1989, Perera had two «Descent…» books).
+
 ## Catalog caching convention
 
 Processed/scraper-unfriendly catalogs are cached under `data/catalogs/` (committed, with source