v0.2.0

• lib verify "quote" [--book ID] [--json] [--limit N]: catches misquoted book text,
  deterministically, no LLM. Book mode (--book) does a full scan of one book; shelf mode
  (no --book) attributes the quote across the library via FTS5 candidates.
• Verdicts exact / close / distorted / not_found / null, a word-level diff against
  the source for close/distorted, and grep-style exit codes with a script discriminator
  (exit 1 with empty stdout = error; exit 1 with output = "checked, not confirmed").
• 6th MCP tool, verify_quote — runs in a worker thread so a full book scan doesn't block
  the stdio server; all 6 tool descriptions are now in English (previously Russian).

v0.1.0

• Deterministic pipeline turning FB2, EPUB, DOCX, HTML, TXT/MD, and text-layer PDF into
  clean, token-counted Markdown chapters — byte-identical output for identical input.
• Quality gate: five metrics (coverage, structure, garbage, encoding, dehyphenation) score
  every book into ok / review / failed, with lib doctor for diagnostics.
• lib find: FTS5 full-text search across chapters, titles, and authors, with RU/EN
  stemming and JSON output.
• Built-in MCP server (lib serve / librarian-cli): 5 read-only tools over the same
  reader core, so Claude can browse, search, and pull chapters under a token budget —
  deterministic RAG with no embeddings, no network, no API keys.
• Offline-vendored o200k_base tokenizer — no network calls at install or runtime.
• CLI: ingest · list · get (spec | --budget) · find · info · doctor · reingest · rm · serve.